Learning The Universe: Cosmological and Astrophysical Parameter Inference With Galaxy Luminosity Functions and Colours

MNRAS 000, 1–27 (2024) Preprint 22 November 2024 Compiled using MNRAS LATEX style file v3.
Learning the Universe: Cosmological and Astrophysical Parameter

Inference with Galaxy Luminosity Functions and Colours
Christopher C. Lovell,1★ Tjitske Starkenburg2,3,4 , Matthew Ho,5 Daniel Anglés-Alcázar,6
Romeel Davé,7,8,9 Austen Gabrielpillai,10 Kartheik G. Iyer,5 Alice E. Matthews,11 William J. Roper,12
Rachel S. Somerville,13 Laura Sommovigo,13 Francisco Villaescusa-Navarro13,14
arXiv:2411.13960v1 [astro-ph.GA] 21 Nov 2024
1 Institute of Cosmology and Gravitation, University of Portsmouth, Burnaby Road, Portsmouth, PO1 3FX, UK
2 Center for Interdisciplinary Exploration and Research in Astrophysics (CIERA), Northwestern University, 1800 Sherman Ave, Evanston IL 60201, USA
3 Department of Physics and Astronomy, Northwestern University, 2145 Sheridan Rd, Evanston IL 60208, USA
4 NSF-Simons AI Institute for the Sky (SkAI), 172 E. Chestnut St., Chicago, IL 60611, USA
5 Columbia Astrophysics Laboratory, Columbia University, 550 West 120th Street, New York, NY 10027, USA
6 Department of Physics, University of Connecticut, 196 Auditorium Road, U-3046, Storrs, CT 06269-3046, USA
7 Institute for Astronomy, Royal Observatory, University of Edinburgh, Edinburgh EH9 3HJ, UK
8 University of the Western Cape, Department of Physics and Astronomy, Bellville, Cape Town 7535, South Africa
9 South African Astronomical Observatories, Observatory, Cape Town 7925, South Africa
10 Department of Astrophysics, The Graduate Center, City University of New York, 365 5th Ave, New York, NY 10016, USA
11 Mullard Space Science Laboratory, University College London, Holmbury St Mary, Dorking, Surrey, RH5 6NT, UK
12 Astronomy Centre, University of Sussex, Falmer, Brighton BN1 9QH, UK
13 Center for Computational Astrophysics, Flatiron Institute, 162 Fifth Avenue, New York, NY, 10010, USA
14 Departent of Astrophysical Sciences, Princeton University, Princeton NJ 08544, USA
Accepted XXX. Received YYY; in original form ZZZ
ABSTRACT
We perform the first direct cosmological and astrophysical parameter inference from the combination of galaxy luminosity
functions and colours using a simulation based inference approach. Using the Synthesizer code we simulate the dust attenuated
ultraviolet–near infrared stellar emission from galaxies in thousands of cosmological hydrodynamic simulations from the
CAMELS suite, including the Swift-EAGLE, Illustris-TNG, Simba & Astrid galaxy formation models. For each galaxy we
calculate the rest-frame luminosity in a number of photometric bands, including the SDSS ugriz and GALEX FUV & NUV
filters; this dataset represents the largest catalogue of synthetic photometry based on hydrodynamic galaxy formation simulations
produced to date, totalling >200 million sources. From these we compile luminosity functions and colour distributions, and
find clear dependencies on both cosmology and feedback. We then perform simulation based (likelihood-free) inference using
these distributions, and obtain constraints on both cosmological and astrophysical parameters. Both colour distributions and
luminosity functions provide complementary information on certain parameters when performing inference. Most interestingly
we achieve constraints on 𝜎8 , describing the clustering of matter. This is attributable to the fact that the photometry encodes the
star formation–metal enrichment history of each galaxy; galaxies in a universe with a higher 𝜎8 tend to form earlier and have
higher metallicities, which leads to redder colours. We find that a model trained on one galaxy formation simulation generalises
poorly when applied to another, and attribute this to differences in the subgrid prescriptions, and lack of flexibility in our emission
modelling. The photometric catalogues are publicly available at: https://camels.readthedocs.io/.
Key words: galaxies: abundances – galaxies: photometry – cosmology: cosmological parameters
1 INTRODUCTION tion of these methods is in Bayesian inference, wherein one estimates

the parameters of a model given some assumed prior hypothesis and
Machine learning methods, particularly deep learning algorithms,
observed data (e.g. Christensen et al. 2001). This combination of
are now capable of learning both deterministic and probabilistic re-
machine learning and Bayesian inference is known as Simulation
lationships with a high degree of flexibility, and have achieved wide
Based Inference (SBI, also called implicit likelihood inference, or
use within the fields of cosmology and astrophysics (Carleo et al.
likelihood-free inference; see Marin et al. 2012; Cranmer et al. 2020),
2019; ?; Smith & Geach 2023). One particularly powerful applica-
a rigorous Bayesian approach to learning the relationship between
simulated data (X) and the underlying parameters used to generate
it (𝜃), typically leveraging flexible neural density estimators (NDEs,
★ E-mail: christopher.lovell@port.ac.uk (CCL)
© 2024 The Authors

2 C. C. Lovell et al.
e.g. Durkan et al. 2019). In contrast with traditional Bayesian meth- The majority of these studies have performed inference on intrin-
ods, an analytic form of the likelihood does not need to be provided; sic properties of galaxies (or physical fields). However, in order to
instead, an SBI approach directly learns the form of the likelihood. make a like-for-like comparison with observations we must forward
This provides many advantages, particularly in instances where the model the electromagnetic emission from galaxies, and compare di-
likelihood is non-Gaussian (such as when systematics are considered; rectly in this observer space (Fortuni et al. 2023). There are a number
Jeffrey et al. 2021), or intractable (such as in field–level inference; of approaches for achieving this, but almost all start by modelling
Leclercq & Heavens 2021; Lemos et al. 2023). They also permit the stellar emission using stellar population synthesis (SPS) models
amortized inference, whereby the full posterior can be evaluated and (Conroy 2013) coupled to the star formation and metal enrichment
sampled rapidly at the point of inference, reducing the computational history of each simulated galaxy. Many also account for line and
cost; this opens up new avenues for Bayesian parameter inference on continuum emission from young nebular regions, utilising photoion-
extremely large datasets (Ho et al. 2024). isation codes such as Mappings and Cloudy (Dopita & Sutherland
A number of studies have used SBI to perform cosmological pa- 1996; Chatzikos et al. 2023). One of the most important galaxy com-
rameter inference. These exploit simulations that explicitly model ponents that has an outsized impact on the observed spectral enbergy
the relationship between cosmological parameters and observable distribution (SED) is cosmic dust, which leads to attenuation in the
statistics, such as the cosmic microwave background power spec- UV–optical, and thermal emission in the infrared (Calzetti 2001;
trum (e.g. Lemos et al. 2023), or clustering statistics in the presence Draine 2003). There are a number of methods for modelling this
of weak-lensing effects (e.g. von Wietersheim-Kramsta et al. 2024). attenuation and emission, from simple analytic models (Charlot &
These simulations cover a wide range of sophistication, from analyt- Fall 2000) to full dust radiative transfer approaches (Camps & Baes
ical descriptions of the distribution of halo properties (Mead et al. 2015; Narayanan et al. 2021).
2021; Tessore et al. 2023) to full 𝑁-body simulations of the large It is clear from these various ingredients in the forward model that
scale structure (Villaescusa-Navarro et al. 2020). Analytic or empir- many degeneracies can be introduced. For example, the choice of SPS
ical models are often then applied to map galaxies or observational model can have a large impact on the predicted intrinsic emission,
tracers to the underlying matter density. However, these simple mod- particularly if the impact of binaries is taken into account (Wilkins
els relating dark matter properties to observational tracers are often et al. 2016; Stanway et al. 2016; Eldridge et al. 2017; Tortorelli et al.
inadequate (Hadzhiyska et al. 2020), and at small scales astrophys- 2024; Bellstedt & Robotham 2024). Photoionisation modelling intro-
ical effects, such as energetic feedback from active galactic nuclei duces a large number of additional parameters, such as the hydrogen
(AGN) and star formation, can impact the overall matter distribution density of the incident cloud, the value of the assumed ionisation pa-
(Anglés-Alcázar et al. 2017; Borrow et al. 2020; Delgado et al. 2023; rameter, as well as the chemical abundances of the cloud and source
Tillman et al. 2023; Gebhardt et al. 2024), requiring modifications to (Byler et al. 2017; Wilkins et al. 2020). The dust attenuation law can
𝑁-body predictions (Bose et al. 2021; Carrilho et al. 2022; Schaller also take a number of forms, dependent on the grain size distribution
et al. 2024a). To overcome these limitations requires simulating these and composition (Steinacker et al. 2013), and depend on the intrinsic
processes explicitly. properties of galaxies in various ways (Salim & Narayanan 2020;
State-of-the-art hydrodynamic simulations are now capable of pro- Trayford et al. 2020; Hahn et al. 2022), such as by scaling the over-
ducing realistic galaxy samples in large volumes that match a num- all attenuation by the mass and metallicity of star forming gas (see
ber of key distribution functions in the local Universe (Somerville & Trayford et al. 2015; Lovell et al. 2019; Vĳayan et al. 2021; Pallottini
Davé 2015; Vogelsberger et al. 2020; Crain & van de Voort 2023). et al. 2022). Despite this flexibility in the forward model, achieving a
These simulations self-consistently model the evolution of collision- match to observations has still proven exceedingly difficult. Optical
less dark matter as well as baryons down to some resolution limit, luminosity functions, well constrained observationally in the local
beyond which subgrid models are used to approximate the impact (𝑧 ∼ 0.1) Universe, have been reproduced with varying success, and
of sub-resolution physical processes on the macro-scale properties accurate colour distributions have proven to be even more elusive;
of galaxies. Such simulations can be used in an SBI framework not achieving the right balance of blue (star-forming) to red (passive,
only to understand the impact of baryons and astrophysical processes dust obscured) galaxies is a complex challenge (Trayford et al. 2015,
on cosmological parameter inference, but also to perform inference 2017; Nelson et al. 2018; Lagos et al. 2019; Donnari et al. 2019;
directly on astrophysical parameters of interest. However, hydrody- Bravo et al. 2020; Dickey et al. 2021; Trčka et al. 2022; Gebek et al.
namic simulations are exceedingly computationally expensive, limit- 2024).
ing the volume, resolution and / or physical fidelity of the simulations Despite these difficulties, a few recent works have forward mod-
that can be run. Beyond the challenges inherent in running the simu- elled observed galaxy properties from hydrodynamic simulations
lations, they are also computationally challenging for use in SBI due and used these within an SBI framework. Recently, Choustikov
to the range of parameters that can be varied, and the large volume et al. (2024) forward modelled JWST NIRCam photometry from
of data products that must be computed and stored. the Sphinx simulations to estimate the ionising emissivity of high-
Despite these obstacles there has been recent progress in producing redshift galaxies. Of particular relevance for this study is the work of
hydrodynamic simulation suites that are amenable to SBI approaches. Hahn et al. (2024), who showed how competitive cosmological con-
The Cosmology and Astrophysics with MachinE-Learning Simula- straints can be achieved using galaxy photometry alone calculated
tions (CAMELS; Villaescusa-Navarro et al. 2021b) project brought from the Illustris-TNG suite of CAMELS; they combined the poste-
together a number of hydrodynamic and semi-analytic galaxy for- riors from individual galaxies to achieve tighter constraints, comple-
mation models, each consisting of large suites of simulations vary- menting the work of Villaescusa-Navarro et al. (2022) &Echeverri-
ing both cosmological and astrophysical paramaters that have been Rojas et al. (2023); Chawak et al. (2024) that performed parameter
used in an SBI framework for parameter inference (e.g. Villaescusa- inference on single galaxies. Inference on other forward modelled
Navarro et al. 2021a, 2022; Nicola et al. 2022; Hassan et al. 2022; properties from CAMELS, such as diffuse gas emission and ab-
Friedman & Hassan 2022; Thiele et al. 2022; Villanueva-Domingo & sorption, has also been demonstrated (?Butler Contreras et al. 2023;
Villaescusa-Navarro 2022; Shao et al. 2022; de Santi et al. 2023b,a; Tillman et al. 2023).
Perez et al. 2023; Jo et al. 2023; Echeverri-Rojas et al. 2023). In this paper we present the first simultaneous constraints of cos-
MNRAS 000, 1–27 (2024)

LtU: Parameter inference on photometry 3
mological and astrophysical parameters using observable galaxy Each simulation contains 2563 dark matter and baryonic (gas
distribution functions, forward modelled from the CAMELS sim- particle) resolution elements, of mass 𝑚 DM = 6.49 × 107 (Ωm −
ulation suite. All photometric catalogues are made publicly avail- Ωb )/0.251 ℎ −1 𝑀⊙ and 𝑚 b = 1.27 × 107 ℎ −1 𝑀⊙ , respectively,
able at https://camels.readthedocs.io/en/latest/data_ within a periodic volume of (25 ℎ −1 ) 3 Mpc3 . The initial conditions
access.html for the community to use. This paper is released as for each simulation were generated using second order Lagrangian
part of the Learning the Universe collaboration,1 which seeks to learn perturbation theory using MUSIC (Hahn & Abel 2011), assuming
the cosmological parameters and initial conditions of the Universe by the same initial power spectrum of gas and dark matter particles at
leveraging Bayesian forward modelling approaches. The hope is that 𝑧 = 127. In all simulations a number of parameters are kept fixed,
by better understanding the link between cosmology, astrophysics, such as Ωb = 0.049, ℎ = 0.6711, 𝑛𝑠 = 0.9624, 𝑀𝜈 = 0.0 eV,
and forward modelled observables, we can build better models for 𝑤 = −1, Ω𝐾 = 0. The differences between simulations are in the
performing this kind of inference. The paper is arranged as follows. In initial random phases, as well as 4 astrophysical parameters (𝐴SN1 ,
Section 2 we describe the CAMELS simulations, including details 𝐴SN2 , 𝐴AGN1 , 𝐴AGN2 ) and 2 cosmological parameters (Ωm , 𝜎8 ). We
on the simulation sets used in this analysis. We describe our for- stress that, due to differences in the subgrid model implementations,
ward model for predicting galaxy emission and extracting rest-frame the astrophysical parameters in each suite have different meanings.
photometry in Section 3. In Section 4 we explore the distribution This is true even in the case where two simulations vary the same
functions from the various CAMELS suites, and the impact of the quantity (e.g. the wind speed), because of the precise implementation
cosmological and astrophysical parameters. We present our results details, but also because different models hold different secondary
in Section 5, as well as introducing the Learning the Universe Im- parameters constant whilst varying this primary variable. The phys-
plicit Likelihood Inference (LtU-ILI) framework for SBI. We discuss ical meaning of each parameter in each galaxy formation model is
our results in Section 6 as well as the difficulties of performing SBI described in Table 1.
directly on observed relations. We summarise our conclusions in Structures are first identified with the Friends-of-Friends algorithm
Section 7. (FOF; Davis et al. 1985) and substructrues with Subfind (Springel
et al. 2001). We treat the Subfind outputs as discrete galaxies, where
the total stellar mass is > 108 M ⊙ (corresponding to at least 5 star
particles). We note that resolution effects due to discrete sampling
2 THE CAMELS SIMULATIONS
of the star formation history can have a large impact on the resulting
The Cosmology and Astrophysics with MachinE Learning Simula- emission, particularly in the UV (Trayford et al. 2015). We will
tions project (CAMELS; Villaescusa-Navarro et al. 2021b) is a suite explore methods for smoothing the recent star formation in future
of over 14000 cosmological hydrodynamic and 𝑁-body simulations, work. In this work we extract the age and metallicity of each star
designed to explore the effect of cosmological and astrophysical particle in each subhalo, and model the emission from each star
parameter choices on structure formation and galaxy evolution, par- particle independently (described in Section 3).
ticularly through their use within machine learning frameworks. In There are a number of different simulation sets within each
this project we use the hydrodynamic simulations from four different CAMELS suite. The Cosmic Variance (CV) set contains 27 simula-
codes: tions with identical (fiducial) parameters, only changing the initial
random number seed, in order to provide a means of assessing the im-
(i) Simba (Davé et al. 2019), run with the Gizmo code (Hopkins pact of cosmic variance. The Latin Hypercube (LH) set contains 1000
2015). simulations where the 6 parameters are varied using a latin hyper-
(ii) IllustrisTNG (Pillepich et al. 2018; Weinberger et al. 2018), cube within the following ranges: Ωm ∈ [0.1, 0.5], 𝜎8 ∈ [0.6, 1.0],
run with the Arepo code (Springel 2010; Weinberger et al. 2019). 𝐴SN1 ∈ [0.25, 4.0], 𝐴SN2 ∈ [0.5, 2.0], 𝐴AGN1 ∈ [0.25, 4.0], and
(iii) Astrid (Bird et al. 2022; Ni et al. 2022), run with the MP- 𝐴AGN2 ∈ [0.5, 2.0] 2 . The 1 Parameter (1P) set contains 25 simu-
Gadget code. The addition of Astrid to the CAMELS suite is de- lations where one parameter at a time is varied around the fiducial
scribed in (Ni et al. 2023), including details of modifications com- parameters. In the 1P and CV sets the fiducial cosmological param-
pared to the fiducial Astrid model. eters are Ωm = 0.3 and 𝜎8 = 0.8, and the fiducial astrophysical
(iv) Swift-EAGLE (Schaye et al. 2015; Crain et al. 2015), run with parameters are defined at 𝐴 = 1.0.
the Swift code (Schaller et al. 2024b) using the Sphenix smoothed The full CAMELS dataset is available online at https://
particle hydrodynamics implementation (Borrow et al. 2022). The camels.readthedocs.io/en/latest/data_access.html, and
addition of Swift-EAGLE to the CAMELS suite is described in Lovell described in Villaescusa-Navarro et al. (2023). We generate pho-
et al. in prep.. tometry for galaxies in all sets (pipeline described in Section 3, and
These models represent a combination of different subgrid models catalogue in Appendix A), and these are also made available online
for gas cooling, chemical enrichment, star formation, stellar feed- at the same address.
back, black hole assembly and AGN feedback. There are also dif-
ferences in the underlying hydrodynamics solvers; Arepo imple-
ments an unstructured Voronoi tesselation, Gizmo uses a meshless 3 FORWARD MODELLING WITH SYNTHESIZER
finite-mass method, and MP-Gadget and Swift utilise Smoothed
Particle Hydrodynamics (SPH). Together these differences lead to In order to model the emission for all galaxies in the CAMELS
significant macro differences between the predicted galaxy proper- simulation suites we employ a forward model for the UV-optical
ties (Villaescusa-Navarro et al. 2021b; Ni et al. 2022), however the spectral energy distribution that takes into account stellar emission
fiducial simulations all reproduce the galaxy stellar mass function (as and the impact of dust attenuation. Importantly, we choose a forward
well as other key scaling relations) at 𝑧 = 0 to reasonable accuracy. model that is agnostic to the specific subgrid implementation in
1 https://learning-the-universe.org/ 2 For Astrid, the 𝐴AGN2 parameter also ranges from 0.25 to 4.0
MNRAS 000, 1–27 (2024)

Table 1. A summary of the physical meaning of the feedback parameters in each of the simulation suites.
Simulation 𝐴SN1 𝐴SN2 𝐴AGN2 𝐴AGN2

Galactic winds: metallicity
Galactic winds: Thermal BH feedback: Thermal BH feedback:
dependence of the stellar
Swift-EAGLE energy per SNII scaling of the Bondi accretion temperature jump of gas
feedback fraction per unit
event [0.25 − 4.0] rate [0.25 − 4.0] particles [0.5 − 2.0]
stellar mass [0.5 − 2.0]
Galactic winds: Kinetic mode BH feedback: Thermal mode BH feedback:
Galactic winds: wind speed
Astrid energy per unit energy per unit BH energy per unit BH
[0.50 − 2.00]
SFR [0.25 − 4.00] accretion [0.25 − 4.00] accretion [0.25 − 4.00]
Galactic winds:
Galactic winds: wind speed QSO & jet mode BH feedback: Jet mode BH feedback: jet speed
Simba mass loading
[0.50 − 2.00] momentum flux [0.25 − 4.00] [0.50 − 2.00]
[0.25 − 4.00]
Galactic winds: Kinetic mode BH feedback: Kinetic mode BH feedback:
Galactic winds: wind speed
Illustris-TNG energy per unit energy per unit BH ejection speed / burstiness
[0.50 − 2.00]
SFR [0.25 − 4.00] accretion [0.25 − 4.00] [0.50 − 2.00]
Figure 1. Spectral energy distribution of an example galaxy from the Illustris-TNG CV set at 𝑧 = 0.1. Shown is the intrinsic emission from young and old
stars, the combined intrinsic emission, as well as the total attenuated emission. Filter transmission curves for some of the key rest-frame filters used in this work
are also shown.
each galaxy formation model, whether that is Illustris-TNG, Swift- on its age, metallicity and initial stellar mass3 . The ‘current’ stellar
EAGLE, Astrid or Simba. This ensures that the results for each mass of a star particle takes into account the recycling of mass due to
model are comparable; the emission of each galaxy is dependent evolved stellar populations; since SPS models implicitly account for
only on its star formation and metal enrichment history, itself derived this as a function of age, the initial stellar mass must be used as input.
from the distribution of star particle properties. The integrated emission of a galaxy is then the sum of the particle
We use Synthesizer (Lovell et al. in prep., Roper et al. in prep.) to contributions. In order to provide a qualitative assessment of the im-
produce synthetic observables, leveraging flexibility and computa- pact of the choice of SPS model, we generate the emission using two
tional efficiency. Below we describe the different ingredients that go different SPS models: BC03 (Bruzual & Charlot 2003) and BPASS
into our Synthesizer forward model, and the resulting luminosity (Eldridge et al. 2017; Stanway & Eldridge 2018). For BC03, we use
and colour distributions. the Padova 2000 tracks, described in Girardi et al. (2000, 2002). For
BPASS we use v3.2, described in Byrne et al. (2022). For both mod-
els we assume a Chabrier (2003) initial mass function (IMF), with a
high mass cut off of 100 M ⊙ , and a low mass cut off of 0.1 M ⊙ . This is
3.1 Stellar population modelling

3 Initial stellar masses are not available for Astrid, so we assume that all star
The intrinsic stellar emission is produced by coupling each star par- particles have an initial mass equal to 1/4 the initial gas element mass (Bird
ticle in a galaxy to a stellar population synthesis (SPS) model, based et al. 2022).
MNRAS 000, 1–27 (2024)

identical to the IMF assumed in each galaxy formation model, which latter has been shown to contribute negligibly to the galaxy attenua-
ensures consistency between the feedback and enrichment subgrid tion curve (Fischera et al. 2003).
prescriptions and the forward modelled emission. In Simba, the properties of dust are predicted self-consistently
Trayford et al. (2015) showed that applying a 3D or 2D aperture using an on-the-fly creation, growth and destruction model (Li et al.
can impact the luminosities of extended, high mass objects, where 2019); in the interests of consistency with the other galaxy formation
up to 40% of the light comes from an extended intracluster compo- models, we ignore these properties, and adopt the model above. In
nent. However, we choose to neglect aperture corrections, for three future work we will explore the impact of our assumed dust model,
reasons. Firstly, the low resolution of the fiducial CAMELS simula- and perform direct inference on these dust parameters, as well as
tions, in terms of mass (∼ 107 M ⊙ per resolution element) and force marginalisation over the effect of dust in cosmological inference. We
resolution, leads to poorly estimated galaxy sizes. Secondly, since will also explore how to link the attenuation to the properties of the
the volume of each simulation is relatively small, we do not simulate galaxy (e.g. the metallicity of star-forming gas) in a model agnostic
any massive clusters or groups; as such, we are not in a mass regime way.
where aperture corrections are important. And thirdly, the precise
aperture corrections will be strongly dependent on the exact survey
configuration and redshift distribution of sources, a detail we do not 3.3 Photometry
consider in this work. As a result, we choose not to apply any 2D or We generate rest-frame photometry from each of our intrinsic and
3D apertures when estimating photometry. dust–attenuated galaxy SEDs in a range of photometric filters. These
We do not model the impact of nebular emission from young star include the GALEX FUV & NUV bands, top-hat filters centred at
forming regions, nor the contribution of active galactic nuclei; we 1500 Å and 2800 Å, SDSS ugriz, UKIRT UKIDSS YJHK, and John-
leave an assessment of the impact of these components on our results son UBVJ bands5 . These filter profiles are shown in Figure 1. We ad-
to future work. We also ignore the effect of the clumpy intergalactic ditionally calculate rest-frame and observer-frame fluxes in the HST
medium (IGM), since we concentrate on rest-frame bands redward ACS F435W, F606W, F775W, F814W, and F850LP bands, the HST
of Lyman-𝛼, and since this effect will not have a large impact below WFC3 F098M, F105W, F110W, F125W, F140W and F160W, and
𝑧 ∼ 3. JWST NIRCam F070W, F090W, F115W, F150W, F200W, F277W,
We adopt the BC03 models as our fiducial SPS model throughout F356W and F444W. We do not use these bands in this study, but make
the manuscript, unless otherwise stated. Figure 1 shows the inte- them publicly available for the community. Formally, the luminosity
grated intrinsic emission in the UV-optical wavelength range from 𝐿 in each band 𝑋 is defined as so,
an example galaxy taken from one of the Illustris-TNG CV set sim- ∫∞
ulations. 𝐿 𝜈 𝑆 𝑋 (𝜈) 𝑑𝜈
𝐿 𝑋 ≡ 0∫ ∞ , (4)
0
𝑆 𝑋 (𝜈) 𝑑𝜈
3.2 Dust model where 𝑆 𝑋 (𝜈) is the transmission at a given frequency for band 𝑋. We
Dust constitutes only around 0.1% of the total baryons in the Universe then convert the calculated luminosities to absolute AB magnitudes,

today, but has an outsize effect on galaxy emission; around 30% of 𝐿 𝜈 / (erg / s / Hz)
all photons today have been reprocessed by dust grains (Bernstein 𝑀𝑋 = −2.5 log10 − 48.6 , (5)
𝑑 = 10pc
et al. 2002). To account for the effect of dust in a way that is agnostic
to the underlying galaxy formation model, we adopt a simple two- where 𝑑 is the distance modulus and 𝐿 𝜈 is the luminosity density.
component dust model that differentially attenuates young and old We do this for all four galaxy formation models over 34 snapshots
stars, analogous to that first presented by Charlot & Fall (2000). The in the redshift range 𝑧 ∈ [0, 6]. The catalogues for each simulation
model varies the dust optical depth 𝜏 as follows, are publicly available at https://camels.readthedocs.io/. For
this study we use these magnitudes to build rest-frame luminosity
𝜏 = 𝜏cloud + 𝜏ISM 𝑡 ⩽ 𝑡 disp (1) functions and colour distributions at a range of redshifts, explored in
𝜏 = 𝜏ISM 𝑡 > 𝑡disp (2) more detail below.
where 𝜏cloud = 0.67, 𝜏ISM = 0.334 and the dispersion time 𝑡disp =
10 Myr. These parameters were calibrated to nearby galaxies, and 4 LUMINOSITY FUNCTIONS AND COLOURS IN CAMELS
have been used in previous studies employing similar models (Genel
et al. 2014; Trayford et al. 2015). The only property of each galaxy 4.1 CV set distributions
(and by extension each galaxy formation model) that directly affects
4.1.1 Galaxy formation model comparison
the level of attenuation is the ratio of stellar mass formed before or
after 𝑡 disp . The Cosmic Variance (CV) set contains 27 simulations, each using
We use a fixed Milky Way attenuation curve (Pei 1992), the same fiducial parameters, but varying the random seed of the
parametrised using the analytic form of Li et al. (2008). The trans- initial conditions. Figure 2 shows the rest-frame luminosity func-
mission at wavelength 𝜆 for a particle with age 𝑡 is then given by tion (LF) from all 27 CV set simulations combined, for each galaxy
formation model at 𝑧 = 0.1. Each LF is normalised by the total com-
𝜏MW (𝜆)
𝑇 (𝜆, 𝑡) = exp − × (𝜏cloud + 𝜏ISM ) (3) bined volume of the simulations (27 × (25 ℎ −1 ) 3 Mpc3 ). We show
𝜏MW (𝜆 = 5500 Å) the GALEX FUV, SDSS 𝑔 and 𝑖 and 2MASS 𝐾 band to demonstrate
We do not model thermal re-emission from dust, nor scattering; the the full wavelength range probed (1500 Å - 20 000 Å). We also show
4 Assuming a hydrogen density 𝑁Hcloud = 1.37 × 1021 cm −2 and 𝑁HISM = 5 Transmission curves obtained from the Spanish Virtual Observatory Filter
6.79 × 1020 cm −2 , for an extinction cross section 𝐶ext
𝑉 = 4.86 × 10 −22 (Draine Profile Service, (Rodrigo et al. 2012; Rodrigo & Solano 2020), http://
2003), all calculated at visible wavelengths, 𝜆𝑉 = 5500 Å, svo2.cab.inta-csic.es/theory/fps/index.php?mode=voservice.
MNRAS 000, 1–27 (2024)

Figure 2. CV set Luminosity functions (LFs; in AB magnitudes) at 𝑧 = 0.1 for the Illustris-TNG, Simba, Astrid and Swift-EAGLE simulation suites in the
GALEX FUV, SDSS 𝑔 and 𝑖, and UKIRT K bands. Each LF is built using galaxies from all CV simulations combined. Each panel shows the LFs obtained from
the attenuated (solid lines) and intrinsic (dashed lines) emission.
Figure 3. The same as Figure 2, but showing normalised colour distributions. An additional magnitude cut for galaxies above 𝑀𝑟 < −20 has been applied to
remove faint galaxies. Left to right: GALEX FUV − NUV, SDSS 𝑔 − 𝑟, 𝑟 − 𝑖, 𝑖 − 𝑧.
both the intrinsic and dust attenuated LFs in each case. It’s clear that colours. Astrid and Simba show similar colour distributions at all
there are significant differences between the different galaxy forma- wavelengths, with a single strong peak in all distributions that corre-
tion models in all bands. Astrid has the lowest normalisation across sponds to a predominantly blue star forming population. Conversely,
the magnitude range in all bands, which matches what is seen for Swift-EAGLE and Illustris-TNG show more bimodal distributions,
the galaxy stellar mass function (Ni et al. 2022). Simba extends to with a more pronounced red population at all wavelengths.
brighter magnitudes in all bands, but most noticeably at the blue
end, where there is almost an order of magnitude more 𝑀FUV ∼ 21
galaxies than the other models. This may reflect the slightly higher 4.1.2 Redshift evolution
normalisation of the galaxy stellas mass function at the high mass end
in Simba (Davé et al. 2019). Both Illustris-TNG and Swift-EAGLE We have also generated rest-frame luminosity functions at a range
show similar behaviour in the form and normalisation at all wave- of redshifts, shown in Figure 4, showing the evolution in abundance
lengths, but the latter tends to have more faint galaxies in all bands, at different wavelengths6 . In all galaxy formation models the FUV
by approximately 0.5 dex. Dust attenuation has a more significant LF rises quickly, peaking at 𝑧 ∼ 2, before falling again by 𝑧 =
impact on bluer bands, as expected from the form of the attenuation 0. This corresponds to the evolution in the cosmic star formation
curve. rate density (Madau & Dickinson 2014); more recent star formation
leads to higher UV emission. Interestingly, the shape of the LF also
We also show a number of UV-optical (normalised) colour dis- evolves with redshift in Simba, showing a more pronounced bright
tributions in Figure 3, including GALEX FUV − NUV and SDSS
𝑔 − 𝑟, 𝑟 − 𝑖 and 𝑖 − 𝑧. These include all galaxies above our fiducial
stellar mass limit of 108 M ⊙ , as well as a further cut to remove 6 We stress that these rest-frame luminosity functions at higher redshift can-
faint galaxies below an 𝑟-band magnitude limit of 𝑀r < −20. In all not be directly compared to observations without accounting for the necessary
cases the inclusion of dust attenuation leads to minor shifts to redder 𝐾-corrections.
MNRAS 000, 1–27 (2024)

Figure 4. Luminosity functions (LFs; in AB magnitudes) in the GALEX FUV, SDSS 𝑔 and 𝑖, and UKIRT K bands, over a range of redshifts (𝑧 ∈ [0, 6]).
Each LF is built using galaxies from all CV simulations combined, normalised by the total volume (27 × (25 ℎ −1 ) 3 Mpc3 ). We show the relations for the
Illustris-TNG, Simba, Astrid and Swift-EAGLE simulation suites. Each panel shows the LFs for both attenuated (solid lines) and intrinsic (faded dashed
lines) emission.
population at intermediate redshifts, before settling to a Schechter- 4.2 1P set distributions

like distribution by 𝑧 = 0. At redder wavelengths this fall in the LF
We now explore the impact of parameter variations on LFs, colours
after the peak of cosmic star formation is less pronounced, and in the
and mass-to-light ratios by studying the simulations in the 1P set,
𝐾 band the evolution mirrors that of the galaxy stellar mass function,
which vary each parameter one by one around the fiducial parameters.
as expected.
Figures 5, 6, 8, 9, 10 and 11 show the GALEX 𝐹𝑈𝑉 and SDSS
𝑟-band LFs, the UV (𝐹𝑈𝑉 − 𝑁𝑈𝑉) and optical (𝑔 − 𝑟) colour distri-
butions, and the FUV and 𝑟-band (subhalo) mass-to-light ratios from
the 1P set for each galaxy formation model at 𝑧 = 0.1. We construct
each mass-to-light (ML) ratio by normalising by the solar value at
MNRAS 000, 1–27 (2024)

Figure 5. 1P set variations of the dust-attenuated photometry at 𝑧 = 0.1 for Illustris-TNG, Simba, Astrid and Swift-EAGLE (columns left to right, respectively),
when changing Ωm . Top row: GALEX FUV luminosity function, with observations from GALEX in the FUV (Budavári et al. 2005). Second row: SDSS 𝑟-band
luminosity function, with observations from the GAMA survey (Loveday et al. 2012). Third row: GALEX FUV-NUV colour distribution, with observational
constraints from GAMA. Fourth row: SDSS 𝑔 − 𝑟 colour distribution, with observational constraints from GAMA. Fifth row: GALEX FUV subhalo mass-to-light
ratio against halo mass (individual objects are plotted where there are fewer than 10 sources in a halo mass bin). Sixth row: binned SDSS 𝑟-band subhalo
mass-to-light
MNRAS 000,ratio1–27against
(2024)halo mass. See Section 6.2 for caveats on the observational constraints.
that wavelength, decrease in the normalisation at the bright end of the UV LF for the
highest Ωm values in these models.
log10 Υ𝑋 / Υ⊙ = log10 (𝑀subhalo / M ⊙ ) / (𝑀𝑋𝐴𝐵 / 𝑀𝑋,⊙
𝐴𝐵
), (6) The colour distributions also tell an interesting story. As Ωm in-
creases, we see redder colour distributions in both the UV and optical.
where 𝑀subhalo is the total subhalo mass, 𝑀𝑋𝐴𝐵 is the AB absolute
𝐴𝐵 is the absolute magnitude of the sun The magnitude of this effect in the optical is largest for the Swift-
magnitude in band 𝑋, and 𝑀𝑋,⊙ EAGLE model; the blue population centred at 𝑔 − 𝑟 = 0.3 for Ωm =
in band 𝑋 (values obtained from Willmer 2018). 0.1 completely disappears when Ωm is increased to 0.5, leading to a
In order to provide a baseline for comparison we also show obser- unimodal distribution of red galaxies (𝑔 − 𝑟 ∼ 0.8). This behaviour
vational constraints from the GAMA survey (Loveday et al. 2012) in suggests that not only is the abundance of haloes and their ML ratios
the 𝑟-band, GALEX in the FUV (Budavári et al. 2005), and the UV changing with Ωm , but also that this effect is wavelength dependent.
(𝐹𝑈𝑉 − 𝑁𝑈𝑉) and optical (𝑔 − 𝑟) rest-frame colours from GAMA Also interesting is that there are less degeneracies between models
(Liske et al. 2015) as compiled by Pandya et al. (2017), all at 𝑧 ∼ 0.1. with high Ωm in the UV and optical colour spaces, with clear differ-
We stress that these observationally derived rest–frame relations are ences in their colour distributions throughout the range of Ωm probed.
not directly comparable to the results from CAMELS presented here, This will become important when performing SBI to infer the value
for reasons described in depth in Section 6.2; here we just mention of Ωm from these distributions, discussed further in Section 5.3.
that these observations are necessarily 𝐾-corrected, and cover dif- Since a higher Ωm leads to not only a higher abundance of haloes
ferent volumes to those in CAMELS, making direct comparisons but also their accelerated formation (McClintock et al. 2019), the
difficult. stellar ages in these massive haloes may be generally higher, lead-
From these figures it is clear that each parameter has a different ing to the reduced abundance of UV-bright galaxies, and an overall
impact depending on the wavelength and galaxy formation model reddening of their colours. Quiescent fractions have not been stud-
considered. We first explore the cosmological parameters, before ied in the fiducial Astrid simulation at 𝑧 = 0.1, but in EAGLE they
moving on to the impact of the astrophysical parameters. show a strong correlation with galaxy mass (Furlong et al. 2015), in
agreement with observational constraints on the passive fraction, and
supporting this narrative. However, similar behaviour is also seen in
4.2.1 Changes in Ωm Illustris-TNG (Donnari et al. 2019, 2021; Gabrielpillai et al. 2022)
An increase in Ωm leads to an increase in the overall normalisa- and Simba (Davé et al. 2019; Rodríguez Montero et al. 2019; Ap-
tion of the halo mass function (Villaescusa-Navarro et al. 2021b), pleby et al. 2020). Simba produces smoother star formation histories
due to accelerated formation of more massive haloes (McClintock that tend to form later (Dickey et al. 2021), which may also explain
et al. 2019). However, it has been unclear how this translates into the elevated recent SFRs at higher Ωm , since any long-term variabil-
changes in the LFs beyond the increase in the halo number density, ity will have had less time to imprint on the overall SFH (Zheng et al.
since the emission is directly dependent on the star formation and 2022).
metal enrichment history of each galaxy, which itself depends in a Mergers between haloes and their host galaxies are also more
complicated way on the underlying cosmology (Iyer et al. 2020, , in frequent in a universe with higher Ωm . Rodríguez Montero et al.
prep.). (2019) showed how mergers have very little effect on passive fractions
We see in Figure 5 that increasing Ωm leads to an increase in the in Simba, suggesting this effect will have little impact on recent star
number of faint galaxies in the optical (𝑟-band), with little depen- formation, and therefore the UV emission, in this model. Dickey et al.
dence on the galaxy formation model, similar to how it impacts the (2021) showed how passive fractions at low-masses are elevated in
galaxy stellar mass function (Ni et al. 2023; Lovell et al. 2023; Jo et al. isolated galaxies in EAGLE compared to Illustris-TNG and Simba,
2023). We see a similar relationship in the UV, though the relations but lower for the most massive objects, which makes it difficult to
are more noisy, which reflects the sensitivity of the UV emission to pick apart the impact of (recent) mergers in these models.
star formation on relatively short timescales (last ∼ 100 Myr; Lee
et al. 2009). In general, however, for values of Ωm > 0.3 the opti-
cal and UV LFs are mostly degenerate, with very little differences
4.2.2 Changes in 𝜎8
across all magnitudes. The UV and optical ML ratios in Figure 5
help explain this behaviour. Whilst the overall normalisation of the The value of 𝜎8 encodes the degree of matter clustering, and it
halo mass function increases with increasing Ωm , the ML ratio also has an interesting, and perhaps unexpected, impact on the LFs and
increases; galaxies in equal mass haloes are fainter in a universe colours. Increasing 𝜎8 does not have a strong effect on the optical
with a higher Ωm . This trend is monotonic for all models, except LFs in all models, nor the optical ML ratios (except a small increase
Swift-EAGLE, which shows a more complicated evolution for Ωm = for increasing Ωm in Simba at halo masses ∼ 1012 M ⊙ ). However,
0.1 (the ML ratio in the UV and optical is much higher in low mass 𝜎8 does have a discernible effect on the UV LF in some models.
haloes). This effect is most pronounced in Simba, where the magnitude of
Above a halo mass of 1011 M ⊙ there is a general trend across the brightest object in the UV increases by up to 1.5 magnitudes to
all simulations and parameters for the ML ratio to decrease with in- 𝑀FUV = −20 for 𝜎8 = 0.6. This can be explained by increases in
creasing subhalo mass; this explains the characteristic schechter-like the UV ML ratios of up to 0.8 dex in Simba as 𝜎8 increases, much
shape of the luminosity function in these bands at faint magnitudes greater than any other model.
(Bullock & Boylan-Kolchin 2017). The mass at which this fall occurs However, the strongest dependence on 𝜎8 is seen in the colours;
is relatively insensitive to Ωm , which explains the lack of correlation increasing 𝜎8 tends to lead to redder colours in the UV, but partic-
of the shape of the LFs with Ωm (except for in Swift-EAGLE). The ularly in the optical (𝑔 − 𝑟), for all models. In Swift-EAGLE and
magnitude of the dependence of the ML ratio on Ωm is also model Illustris-TNG this manifests as a relative shift in the height of the red
dependent; the largest variations at a halo mass of 1012 M ⊙ are seen and blue populations in their respective bimodal colour distributions.
in Swift-EAGLE and Astrid, with a difference in the ML ratio in In contrast, in Simba and Astrid this manifests as a shift in the peak of
the UV of |Υ𝑋 | > 2. This explains why there is also a significant the unimodal distribution. This suggests that there is a fundamental
MNRAS 000, 1–27 (2024)

Figure 6. The same as Figure 5, but showing the variation in the 𝜎8 1P set.
MNRAS 000, 1–27 (2024)

difference in how changes in 𝜎8 manifest in differences in the colour man et al. (2023) show how increasing 𝐴SN1 leads to a decrease in
distributions. the overall number density of black holes, reducing the global impact
Since the emission is wholly dependent on the star formation and of AGN feedback. This latter effect also has an obvious impact on
metal enrichment history (SFZH) of our galaxies, we explore how the colours, leading to a much bluer distribution in the UV-optical.
this SFZH, and its marginal distributions, is dependent on the value It may also reflect that increasing 𝐴SN1 leads to greater gas heating,
of 𝜎8 to try and understand these correlations. Figure 7 shows the rather than expelling gas from haloes, so that gas is still available for
normalised star formation history, using the summed ages of all star star formation at later times once it has cooled.
particles in each volume, from the Illustris-TNG 1P set changing On the contrary, 𝐴SN2 , which controls the speed of galactic winds,
𝜎8 , further binned by stellar metallicity. It is clear that, for higher has a large impact on the bright end of each LF. This is reflected in the
𝜎8 , galaxies form earlier, and the aggregate stellar population age is ML ratios, where there is a strong dependence on 𝐴SN2 in relatively
therefore older. We find a similar trend for the other models. This is higher mass haloes (> 1012 M ⊙ ) This may reflect the fact that gas has
the primary driver of the redder colours seen in Figure 6. been fully expelled from haloes by the higher wind speeds, reducing
Figure 7 also shows the normalised metallicity distributions as a the overall gas fractions (though the details of baryon spread can
function of 𝜎8 in Illustris-TNG, in bins of stellar age. These show a depend in a complex way on feedback parameters, see Gebhardt
more subtle dependence on 𝜎8 ; greater clustering leads to marginally et al. 2024). Impacts on the growth and accretion onto nearby massive
higher metallicites, but only in the oldest stellar populations. This haloes, containing the brightest galaxies, may also be a secondary
may have a second order impact on the colour distributions, but effect. For 𝐴SN2 the secondary effects on the black hole number
the main driver is the stellar ages. A more detailed exploration of density are far less prominent.
the impact of cosmological and astrophysical parameters on the star In Simba, 𝐴SN1 controls the mass loading of stellar winds, and
formation and metal enrichment histories of galaxies in CAMELS is 𝐴SN2 controls the wind speed. However, contrary to Illustris-TNG,
provided in Iyer et al. in prep.. increasing the wind speed (𝐴SN2 ) leads to brighter galaxies in the UV
For larger 𝜎8 , galaxies of all masses will tend to live in denser for higher wind speeds. The ML ratios reflect this, showing a turnover
environments, closer to massive haloes and filaments, which may in the 𝐴SN1 dependence above 1012 M ⊙ . Gebhardt et al. (2024)
have an additional impact on galaxy colours. Bulichi et al. (2024) showed how increased wind speed actually leads to reduced baryon
explored how this impacts galaxy properties in the Simba model; spread, since less gas is available in the central regions for AGN
they found that SFRs tend to be reduced with increasing proximity accretion, which may explain why there is increased star formation
to massive neighbours, that satellites are affected more than centrals, leading to higher UV emission. Tillman et al. (2023) showed that
and that increased shock heating in large scale structures such as fila- 𝐴SN2 in Simba also curtails SMBH growth, which would contribute
ments, rather than the impact of feedback, drives these reductions in to the bright end behaviour of the LFs, and also the bluer optical
SFR. Bulichi et al. (2024) also compared their results for Simba with colour distributions. Also counter intuitively, increasing the mass
EAGLE and Illustris, and found that the trend of SFR with filament loading of winds (𝐴SN1 ) leads to an increase in the number of UV
proximity is smoother in Simba, compared to a more abrupt change bright galaxies, and a decrease in the low-mass normalisation, which
in EAGLE and Illustris-TNG. These effects may explain the bi-modal is also reflected in the increased fraction of blue galaxies.
colour distributions seen in these latter models, and contribute to the In Astrid, 𝐴SN1 controls the feedback energy per unit SFR in
relative impact of changes in 𝜎8 on the colour distributions. stellar winds, and 𝐴SN2 the wind speed, both similarly to Illustris-
TNG. 𝐴SN1 slightly reduces the faint-end abundance in the optical,
due to increased ML ratios for higher 𝐴SN1 below 1012 M ⊙ . In the
4.2.3 Changes in supernovae feedback parameters UV it is difficult to discern any clear trends due to the low number
The 𝐴SN1 and 𝐴SN2 parameters control different elements of the of galaxies, however the UV ML ratios show the opposite trend to
subgrid star formation feedback models in each galaxy formation the optical above 1012 M ⊙ , with higher ML ratios for lower 𝐴SN1 .
model (see Table 1 for a summary), and as such it is not possible asntwo has a very strong impact on the global normalisation of the
to compare them directly. Instead we explore each galaxy formation LF in the UV and optical, leading to a drop of over 1 dex across
model in turn, and discuss the impact of each parameter change on all magnitudes. The implementation of the wind speed affects both
that specific model. Figure 8 show the variation of the LFs, colour high and low mass haloes, as clearly seen in the ML ratios, reducing
distributions and ML ratios for 𝐴SN1 , and Figure 9 shows the same star formation and the relative abundance across all halo masses.
for 𝐴SN2 . As a result, the number of galaxies in Astrid that reach the stellar
In Illustris-TNG, 𝐴SN1 controls the feedback energy per unit SFR mass and 𝑟-band magnitude brightness limits for the colour plots in
in the form of stellar winds, whereas 𝐴SN2 controls the speed of Figure 9 is significantly reduced for high 𝐴SN2 , leading to very noisy
those winds. Both parameters have a significant impact on the UV distributions; despite this the overall trends are for bluer distributions
and optical LFs. In the optical, increasing 𝐴SN1 tends to lead to a at higher feedback parameters.
reduction in the normalisation at all magnitudes, which is reflected Finally, in Swift-EAGLE 𝐴SN1 controls the energy per unit SNII
in the increasing normalisation of the ML ratio at all halo masses feedback event, which drive galactic winds, and 𝐴SN2 controls the
with increasing 𝐴SN1 . This increase is greatest for haloes with mass metallicity dependence of the feedback fraction per unit mass, 𝑛 𝑍 ,
between 1011 − 1012 M ⊙ , which reflects the fact that feedback from 𝑓th,max − 𝑓th,min
𝑓th = 𝑓th,min + , (7)
stellar winds affects these relatively lower mass haloes more effec- 𝑛H,birth −𝑛𝑛
𝑛𝑍
𝑍
tively. In the UV, the evolution is more complex; the UV ML ratio 1+
0.1 𝑍 ⊙ 𝑛H,0
increases with increasing 𝐴SN1 at low halo masses (< 1012 M ⊙ ), but
above this mass limit the relation reverses. This leads to a reduction where 𝑓th,max and 𝑓th,min are the upper and lower limits of the feed-
in the normalisation of the faint end of the UV LF, but an increase in back fraction, and the (𝑛H,birth /𝑛H,0 ) −𝑛𝑛 term controls the density
the normalisation at the bright end. dependence of the feedback fraction. Increasing 𝐴SN1 leads to an
This behaviour may be explained by the secondary effect of stellar overall reduction in the UV and optical LF normalisation. In the op-
feedback suppressing supermassive black hole (SMBH) growth; Till- tical this is clearly explained by the monotonic relationship between
MNRAS 000, 1–27 (2024)

Figure 7. Top: normalised cosmic star formation history in bins of stellar metallicity, from the Illustris-TNG 1P set, changing 𝜎8 . Bottom: normalised stellar
metallicity distribution in bins of stellar age.
ML ratio and 𝐴SN1 , however in the UV things are more compli- In Astrid 𝐴AGN1 controls energy per unit SMBH accretion, and
cated, with clear dependencies of UV ML on halo mass that further this particular parameter also has no impact on the LFs and colours.
depend on the value of 𝐴SN1 . As for the colours, increasing 𝐴SN1 However, 𝐴AGN2 , which controls the energy injected per unit accre-
leads to bluer galaxy distributions; this may reflect secondary effects tion mass in the thermal feedback mode, has a large impact on the
on the growth of SMBHs, as discussed previously, or the fact that LFs, leading to an increase in excess of 1.5 dex at the bright end of
gas is heated rather than expelled, meaning there is fuel for later the UV LF, and 1 dex in the optical. Why does increasing this par-
star formation. Increasing 𝐴SN2 leads to increased ML ratios in the ticular channel of BH feedback lead to increased number densities?
optical and UV, and subsequently reduced normalisation of the LFs The reason is efficient self-regulation: enhanced thermal feedback
at all magnitudes. It also leads to redder distributions, which reflects severely suppresses the growth of SMBHs, reducing the overall feed-
the fact that galaxies that enrich early are then susceptible to more back injected (particularly in the jet mode), leading to enhanced star
powerful feedback at later times, reducing their star formation. formation (Ni et al. 2023). This also leads to bluer colours in both the
UV and optical. Another reason why we see a stronger effect from
the AGN parameters in Astrid compared to Simba and Illustris-TNG
4.2.4 Changes in AGN feedback parameters is that Astrid has a higher number density of massive black holes in
the fiducial model, as evidenced by the 0.5 dex higher normalisation
As is the case for the supernovae feedback parameters, 𝐴AGN1 and of the black hole mass function at the high-mass end (Ni et al. 2023).
𝐴AGN2 control different elements of the subgrid AGN feedback mod-
els in each galaxy formation model (see Table 1 for a summary), and
cannot be compared directly. Contrary to the supernovae param-
eters, the AGN parameters have a much reduced effect in almost Finally, in Swift-EAGLE AGN feedback is implemented in a
all the galaxy formation models. This reflects the small box size purely thermal mode; 𝐴AGN1 controls the efficiency scaling of the
in CAMELS ((25 ℎ −1 ) 3 Mpc3 ) and subsequent lack of (relatively) Bondi accretion rate, and 𝐴AGN2 the temperature jump of the gas
massive haloes, as well as SMBH self-regulation. during AGN feedback events. Neither parameter has a large effect
In Illustris-TNG, 𝐴AGN1 controls the kinetic energy released per on the optical LFs, but there is a small difference in the bright end
unit SMBH accretion mass, and 𝐴AGN2 the ejection speed and bursti- in the UV, whereby increasing both parameters reduces the abun-
ness of that ejected material (where greater burstiness leads to more dance of UV bright galaxies. The differences in the colours are more
frequent but lower energy feedback events; Tillman et al. 2023). How- pronounced; increasing both parameters lead to redder UV and op-
ever, neither has any appreciable effect in either the LFs or colour tical colours. This suggests that increasing the primary efficacy of
distributions above and beyond butterfly effects (Genel et al. 2019). thermal feedback in Swift-EAGLE does not lead to the same self-
A similar story unfolds for Simba; in this model 𝐴AGN1 controls the regulation effects seen in Astrid, and had a more predictable effect
momentum in quasar and jet mode feedback, and 𝐴AGN2 the speed on the luminosities and colours of massive, bright galaxies hosting
of the jet. However, neither has any impact on the LFs and colours. AGN.
MNRAS 000, 1–27 (2024)

Figure 8. The same as Figure 5, but showing the variation in the 𝐴SN1 1P set.
MNRAS 000, 1–27 (2024)

Figure 9. The same as Figure 5, but showing the variation in the 𝐴SN2 1P set.
MNRAS 000, 1–27 (2024)

Figure 10. The same as Figure 5, but showing the variation in the 𝐴AGN1 1P set.
MNRAS 000, 1–27 (2024)

Figure 11. The same as Figure 5, but showing the variation in the 𝐴AGN2 1P set.
MNRAS 000, 1–27 (2024)

5 SIMULATION BASED INFERENCE WITH LTU-ILI Table 2. Bright- and faint–end AB magnitude limits chosen for luminosity
functions used in the SBI analysis.
We now explore the use of our forward modelled UV-optical lumi-
nosity and colour distributions from the CAMELS LH set within Band Bright limit Faint limit
an SBI framework. SBI is a powerful framework for performing in- FUV -20.5 -15
ference in this domain since it permits amortized posteriors, which NUV -20.5 -15
significantly speeds up the process of obtaining posteriors. This al- u -21.5 -16
lows us to apply the trained model to hundreds of test set simulations g -22.5 -17
in just fractions of a second. Additionally, SBI does not require an r -23.5 -18
i -24 -18.5
explicit likelihood, which permit us to use more complex noise mod-
z -24.5 -19
els, which we will explore in future work when comparing directly
to observations.
Table 3. Blue- and red–limiting colours chosen for the construction of colour
distributions used in the SBI analysis.
5.1 SBI Architecture
Colour Bright limit Faint limit
We use the LtU-ILI package (Ho et al. 2024) to perform SBI. LtU-ILI 𝐹𝑈𝑉 – 𝑁𝑈𝑉 -0.5 3.5
is a codebase that simplifies many of the steps in an SBI analysis, par- 𝑔–𝑟 0.0 1.0
ticularly those concerning testing and validation such as estimating 𝑟 –𝑖 0.0 0.5
the epistemic uncertainty, choosing hyperparameters, and assessing 𝑖–𝑧 -0.1 0.4
model misspecification. We use the sbi package within LtU-ILI, and
perform neural posterior estimation (NPE; Papamakarios & Murray all luminosity functions and colour distributions described in Sec-
2016; Greenberg et al. 2019), to allow amortized inference. We con- tion 5.1 at a single redshift, 𝑧 = 0.1. We also calculate a number
struct an ensemble of two models, each composed of a neural spline of statistics describing the goodness of fit, including the Root Mean
flow (Durkan et al. 2019) with 4 transforms, 10 spline bins, and Squared Error (RMSE),
60 hidden features. During training we use a batch size of 4, and a v
u
t 𝑁
fixed learning rate of 5 × 10 −4 , and an improvement based stopping 1 ∑︁
RMSE = (𝜃 𝑖 − 𝜇𝑖 ) 2 , (8)
criterion. 𝑁
𝑖=1
Throughout the rest of this manuscript we will present models
trained and applied to different simulations. We use the same ar- and the mean relative error,
chitecture for each trained model. For each simulation we reserve 𝑁
1 ∑︁ |𝜃 𝑖 − 𝜇𝑖 |
a subset of the LH simulations as a test set (10%), and a further 𝜖= , (9)
𝑁 𝜇𝑖
subset of the remaining simulations for validation (10%). The re- 𝑖=1
maining simulations we use for training, and evaluate our model which both measure the precision of the estimate (lower values being
performance during training on the validation set. Our priors are set more precise). We also calculate the coefficient of determination,
by the chosen distribution of the parameters in the LH set, described Í𝑁
in Villaescusa-Navarro et al. (2021b). After training, to produce pos- (𝜃 𝑖 − 𝜇𝑖 ) 2
𝑅 2 = 1 − Í𝑖=1 , (10)
teriors for a given set of parameters we directly take 1000 samples 𝑁
𝑖=1 (𝜃 𝑖 − 𝜃 𝑖 )
ˆ 2
from the NPE.
For our input data we choose a subset of attenuated luminosity which measures the accuracy of the estimate (values close to 1 being
functions (in the GALEX 𝐹𝑈𝑉 and 𝑁𝑈𝑉 bands, and the SDSS more accurate), and the reduced chi-squared,
𝑢𝑔𝑟𝑖𝑧 bands) and colour distributions (𝐹𝑈𝑉 − 𝑁𝑈𝑉, 𝑔 − 𝑟, 𝑟 − 𝑖, 𝑁
1 ∑︁ (𝜃 𝑖 − 𝜇𝑖 ) 2

𝑖 − 𝑧). For each luminosity function we apply 12 bins linearly spaced 𝜒2 = , (11)
between bright- and faint–end AB magnitude limits given in Ta- 𝑁 𝜎𝑖
𝑖=1
ble 2. For colour distributions, we also apply 12 bins linearly spaced which measures the accuracy of the estimated posterior uncertain-
between the red- and blue–limits given in Table 3. We provide the ties; values close to 1 indicate that the magnitude of the errors (ap-
observational data as a single one-dimensional concatenated vector, proximated by the marginal standard deviation 𝜎) has been estimated
X; where multiple redshifts are considered these are simply concate- correctly, whereas values above / below 1 indicate an under- / over-
nated together. We perform coverage tests for each trained model, prediction of the error. For all galaxy formation models and parame-
discussed in Appendix B. In the following sections we will explore ters (except for a couple of exceptions discussed below) the reduced
how our results depend on the choice of luminosity functions or chi-squared value lies between 0.76 and 1.55, which indicates that
colours, and how combining distributions at different redshifts can all errors are well estimated.
improve constraints on key cosmological and astrophysical parame- There are some clearly visible trends in Figures 12 & 13. Ωm is pre-
ters. dicted precisely and accurately across all simulations, with a RMSE
of 0.03 and 𝑅 2 of 0.93 in Swift-EAGLE, though slightly less well in
Astrid (RMSE=0.07, 𝑅 2 =0.63). This agrees well with the large vari-
5.2 Inference
ations in the LFs and colour distributions as a function of Ωm seen in
We begin by showing the predicted marginal posteriors on our cos- Section 4.2.1. The value of 𝜎8 is also predicted, somewhat surpris-
mological and astrophysical parameters for all test set simulations, ingly; all we have provided as observed data are luminosity functions
using a model trained and tested on the same galaxy formation model, and colour distributions, which do not include any explicit spatial in-
in Figures 12 & 13. We show posterior predictions for Ωm , 𝜎8 , 𝐴SN1 , formation from which to derive clustering constraints. The precision
𝐴SN2 , 𝐴AGN1 and 𝐴AGN2 against their true values, for Illustris- and accuracy strongly depend on the galaxy formation model, with
TNG, Simba, Swift-EAGLE and Astrid. As input data we provide the best precision and accuracy found in Illustris-TNG (RMSE=0.05,
MNRAS 000, 1–27 (2024)

Figure 12. Predicted marginal posteriors on the test set vs the true values for each cosmological (Ωm , 𝜎8 ) and astrophysical (𝐴SN1 , 𝐴SN2 , 𝐴AGN1 , 𝐴AGN2 )
parameter. These estimates use both colours and luminosity functions at 𝑧 = 0.1, described in more detail in Section 5.2, using training and testing data from
the same galaxy formation model; we discuss generalisability across different models in Section 5.4. Top two rows: Swift-EAGLE. Bottom two rows: Astrid.
MNRAS 000, 1–27 (2024)

Figure 13. The same as Figure 12, but showing predictions for Illustris-TNG (top two rows) and Simba (bottom two rows).
MNRAS 000, 1–27 (2024)

𝑅 2 =0.81), and the worst of both in Astrid (RMSE=0.09, 𝑅 2 =0.38). error estimates across the test set. Colours, on the other hand, do
This suggests that specifics of the galaxy formation model affect how lead to very tight constraints on Ωm ; the variation can be seen across
𝜎8 impacts the population galaxy emission, as seen in Section 4.2.2, the range of Ωm in Figure 5, which explains the low errors across
highlighting the importance of taking into account such specifics the whole test set. For 𝜎8 it appears that the use of LFs and colours
when marginalising over astrophysics in a cosmological analyses individually give relatively similar constraints, but combined they
using such observable features. provide much tighter constraints. For 𝐴SN1 and 𝐴SN2 it is a similar
As discussed in Sections 4.2.3 & 4.2.4, the feedback parame- story; comparable constraints for LFs and colours individually, but
ters are not directly comparable between galaxy formation models. tighter constraints combined.
However, in general, the stellar feedback parameters are all con- The fact that colours drive the tight constraints on Ωm is a some-
strained to some degree across all models. Certain parameters, such as what surprising result. Jo et al. (2023) show how using just the
𝐴SN2 for Illustris-TNG, are predicted remarkably well (RMSE=0.09, binned galaxy stellar mass function can achieve tight constraints on
𝑅 2 =0.96), driven by the clear impact on the UV and optical LFs cosmological and astrophysical parameters, even in the presence of
and colour distributions. In contrast, the AGN parameters are, in simulation uncertainty. We do not include the 𝐾-band in this anal-
general, not constrained at all, tending to just predict the median ysis, which is known to closely trace the underlying stellar mass,
of each distribution. There are some exceptions, such as 𝐴AGN1 however one might naively expect the optical luminosity functions to
for Swift-EAGLE (RMSE=0.54, 𝑅 2 =0.66) and 𝐴AGN2 for Astrid still provide sufficient information on the total stellar mass content.
(RMSE=0.76, 𝑅 2 =0.44). The fact that these both control the thermal This suggests that the impact of age and metallicity degeneracies may
feedback mode suggests that this feedback mode may have more of play an important role. It also highlights how the colours trace the
an impact on the stellar contents of relatively lower mass haloes, stellar age of the global galaxy population, which can be significantly
present in these CAMELS simulation volumes. However, the dis- higher in a universe with a lower Ωm (up to 18 Gyr for Ωm =0.1),
cussion in Section 4.2.4 clearly shows how these parameters affect since low mass stars have exceedingly long lifetimes (exceeding 100
galaxy growth, and the resultant observables, in opposite directions; Gyrs; Laughlin et al. 1997).
in Swift-EAGLE, increasing the efficacy of the feedback by dump-
ing more energy inhibits galaxy growth more efficiently, whereas
in Astrid the opposite is true, as increasing the thermal energy per 5.3.2 Redshift Dependence
unit accreted mass curtails SMBH growth, preventing SMBHs from
Figure 14 also shows two example corner plots (right column) when
growing massive enough to start the more effective jet feedback chan-
using the combination of the LFs and colours obtained at different
nel. This may also explain why, despite the large differences seen in
redshifts. In all these examples the marginal posteriors are tighter
the LFs and colours for Astrid, there are significant degeneracies be-
when combining information from different redshifts, as expected.
tween parameters, which degrades the posterior prediction accuracy
We summarise results from the whole test set in Figure 15, as we
compared to e.g. 𝐴SN2 .
did for the LFs and colours. For Ωm , using data at 𝑧 = 0.1 has the most
constraining power, and this degrades at higher redshift. However, the
improvement from combining information from the three redshifts
5.3 Feature Importance and Redshift Dependence
is significant, decreasing the RMSE by >0.01. Interestingly, for 𝜎8
Given the promising constraints on astrophysical and cosmological the most constraining power is provided by the distribution functions
parameters demonstrated in Section 5.2, we now ask what particular at 𝑧 = 2; this suggests that the impact of matter clustering on the
information in our modelled data (or features) is providing con- luminosities and colours of galaxy populations may be more evident
straints on these parameters. We also explore how using information at cosmic noon.
at different redshifts affects our inference, and whether combining It is worth noting that, in these examples, by combining snapshots
distribution functions from multiple redshifts can improve parameter at different redshifts we are at risk of using the same galaxies at
constraints. different points in their evolution. This may bias our constraints
compared to a real survey, since galaxies at different redshifts along
the lightcone will not be associated in the real Universe, and this
5.3.1 Feature Importance evolution encoded at different times may be more constraining than
independent measurements. In order to overcome this issue we could
To begin with we show four examples of the posterior constraints on
subdivide each simulation volume and take galaxies from subboxes at
simulations from the Illustris-TNG test set at 𝑧 = 0.1 as 2D corner
different redshifts, however for the CAMELS boxes considered here
plots (Figure 14). The left column shows the constraints achieved
the volumes are too small to get sufficient statistics. We will more
when providing luminosity functions (LFs) only, colour distributions
robustly assess the impact of these correlations in future work with
only, or the combination of LFs and colours. These clearly show
the larger volume CAMELS-SAM simulations (Perez et al. 2023).
the impact of the different features on the constraints for different
parameters in these specific examples, and the correlations between
the constraints on different parameters in some cases.
5.4 Generalization Across Different Models
To better summarise the impact of the different features across the
whole test set, we combine the posteriors in Figure 15 by binning the One of the motivations for including multiple galaxy formation sim-
stacked residuals; where these are more peaked at zero, the tighter the ulations in the CAMELS suite was to enable tests of the robustness
constraints on this parameter. Here we can clearly see the importance of trained machine learning algorithms to different training simu-
of the various features on each parameter. Interestingly, it appears lations (de Santi et al. 2023b; Ni et al. 2023). In this case, we can
that LFs do not place tight constraints on Ωm . The variation with use the forward modelled photometry from one simulation as input
Ωm seen in Figure 5 in the FUV and 𝑟-band is mostly for low values to our SBI pipeline trained on another, and test the recovery of the
of Ωm (< 0.3); at higher values the differences in the LFs are much underlying parameters. Since the astrophysical parameters represent
smaller, which introduces large degeneracies, and leads to inflated different aspects of each model we cannot perform inference on these
MNRAS 000, 1–27 (2024)

Figure 14. Posterior 2D and 1D marginal posteriors for two randomly selected test set simulations from the Illustris-TNG suite. Each column of corner plots
shows the same test set simulation posteriors, assuming different input features. The true values are indicated by the black dashed lines. Top row: posteriors
given luminosity functions (blue), colours (orange), or both combined (green). Bottom row: posteriors given luminosity functions and colours at 𝑧 = 0.1 (navy),
𝑧 = 2.0 (olive), and both redshifts combined (pink).
between models, so instead we focus on the cosmological parameters or just colour distributions, as well as reducing the fidelity of the
Ωm and 𝜎8 . distribution functions by reducing the number of bins, and achieve
similarly poor results.
We find very poor recovery of parameters when testing between
simulations, and this is the case across all simulations used for train- The main source of this lack of robustness is the very different sub-
ing the SBI framework and testing. The reason for this can be clearly grid prescriptions in each model, which lead to different distributions
seen in Figures 5, 6, 8, 9, 10 and 11, which show the range of lumi- of point-in-time properties, such as stellar mass and star formation
nosity functions and colour distributions across the 1P sets of each rate, as well as different overall star formation histories (Iyer et al.
subgrid model. There is very little agreement or overlap in the de- in prep.). Another source of inflexibility is in our forward model for
tailed distributions, which leads to out-of-distribution errors when galaxy emission, which assumes a simple dust prescription, and fixes
performing inference. We have tested using just luminosity functions many key parameters, such as the nebular cloud dispersion time. In
MNRAS 000, 1–27 (2024)

Figure 15. Combined marginal posterior distribution of Ωm , 𝜎8 , 𝐴SN1 and 𝐴SN2 from all test set objects in the Illustris-TNG suite. Left: models trained using
luminosity functions (blue), colours (orange), or both combined (green). Right: models trained using luminosity functions and colours combined at 𝑧 = 0.1
(navy), 𝑧 = 1.0 (green), 𝑧 = 2.0 (olive), and all three redshifts combined (pink).
future work we will self-consistently modify elements and parame- by 𝜎8 , are another surprising result presented in Section 5.2. For all
ters of the forward model, such as the dust attenuation model, which four simulations we achieve a RMSE ⩽ 0.09, despite not including
should lead to increased overlap in colour and luminosity space be- any spatial information in our features at all. In Section 5.3 we show
tween models. We will also explore contrastive learning (Le-Khac that it is a combination of galaxy colours and luminosity functions
et al. 2020) and domain adaptation approaches (Roncoli et al. 2023; that drives this good predictive accuracy in Illustris-TNG, which
Ćiprĳanović et al. 2020, 2022), which have shown promise in over- is in turn driven by the older ages of galaxy stellar populations in
coming these issues when combining models, and for extracting simulations with a higher 𝜎8 .
domain-invariant features, as well as approaches for excluding sig- We stress that when interpreting these constraints quantitatively
nificant outliers (Echeverri-Rojas et al. 2023; de Santi et al. 2023a). we must keep in mind that this analysis does not include any obser-
vational uncertainties, which will degrade our constraints (see e.g.
Hernández-Martínez et al. 2024). We discuss other observational
considerations below. Additionally, we do not include any estimate
6 DISCUSSION of the simulation uncertainty; Jo et al. (2023) show that this can be
6.1 Cosmological Constraints derived from the CV set and used within an emulator to assess its im-
pact, and find that it can lead to wider estimated posteriors. Finally,
The constraints on cosmological parameters demonstrated in Sec- our dust model does not depend in any way on galaxy properties.
tion 5.2 are surprising. By just considering UV-optical luminosity Including the correlations and covariances between, for example, the
functions and colours we can obtain precise and accurate estimates mass of dust and the optical depth, or the star-dust geometry and the
of Ωm across different subgrid galaxy evolution models. For Illustris- form of the attenuation curve, will lead to degeneracies with some
TNG, Swift-EAGLE and Simba we obtain a RMSE across the test set of the effects on the LFs and colours seen in Figures 5 and 6, which
of 0.03, and for Astrid slightly higher at 0.07, which represent highly will degrade the constraints on cosmological parameters. However,
significant constraints. These are tighter than those obtained by Hahn conversely, the dust model itself may leak cosmological information,
et al. (2024) on galaxies from the NASA-Sloan Atlas, despite only due to dependencies on the SFZH from each galaxy; we explore this
using summary statistics measured on a fraction of the number of effect with our current galaxy-independent dust model briefly in Sec-
galaxies. In Section 5.3 we show that, in Illustris-TNG, it is pre- tion C. We will explore a more sophisticated model that takes these
dominantly galaxy colours that drive these significant constraints. correlations into account in future work.
This suggests that it is internal galaxy properties that are affected
by changes in the matter density, not just the overall abundance of
haloes and therefore galaxies. This is supported by the dependence
6.2 Performing SBI directly on observations
of the ML ratios on Ωm , which implicitly factors out the change in
halo abundance. These findings somewhat support what was found in When performing inference in a simulation based inference frame-
Villaescusa-Navarro et al. (2022) and Echeverri-Rojas et al. (2023), work on cosmological parameters in particular, one must take into
where the properties of a single galaxy were used to obtain significant account the impact of changing the cosmological parameter on any
constraints on Ωm . The constraints on matter clustering, represented forward modelled properties or distribution functions. In this study,
MNRAS 000, 1–27 (2024)

where we focus on volume normalised rest-frame distribution func- solved short timescale variations in the star formation history. These
tions, this manifests in two main ways; 𝐾-corrections for rest-frame are important for predicting the recent star formation. which is key
photometry, and cosmological volume normalisation. for self-consistently producing the contribution from nebular line and
For rest-frame photometry, the change in luminosity distance im- continuum emission. This recent star formation can be estimated by
plied by a change in Ωm leads to differences in the redshift, and smoothing the star formation history, or even by adding small scale
therefore a change in the rest-frame part of the spectrum probed by variation through power spectral density methods (Iyer et al. 2020),
the photometric filter (at a given redshift). In observed catalogues, however these typically require training from higher resolution in-
the emission within a filter bandpass in the rest-frame is estimated by put simulations. Finally, the application of detailed dust radiative
assuming a 𝐾-correction (Hogg et al. 2002). This correction assumes transfer approaches, whilst also limited by their computational com-
a shape for the spectrum, which depends on the assumed galaxy pop- plexity, are also inappropriate for such low-resolution simulations,
ulation, and its evolutionary history. These properties are explicitly since the star-dust distribution will be poorly estimated. In summary,
cosmology dependent; if the cosmology is updated, the 𝐾-correction whilst the fidelity of the CAMELS simulations is sufficient for an
(or, more specifically, the 𝐸-correction) must be updated in line with an analysis of broad band photometric emission in the UV-optical,
this cosmology. In an SBI framework, updating the target observable higher resolution simulations will be required to explore these more
on each parameter iteration is not feasible. Instead, one could for- detailed spatial and time-sensitive features, and for enabling more
ward model into observer-frame fluxes, to avoid the application of a sophisticated forward modelling approaches.
𝐾-correction altogether.
Similarly, the change in apparent distance implied by a change
in Ωm also leads to changes in the differential comoving volume
at a given redshift. This impacts the volume normalisation of any 7 CONCLUSIONS
derived number count distributions, such as luminosity functions. In
We have forward modelled the UV-NIR emission from galaxies
order to perform a like-for-like comparison, it is necessary to correct
in the CAMELS simulation suites, and used these to perform
for this volume normalisation offset as a function of cosmology.
Simulation Based Inference (SBI) on galaxy luminosity functions
Alternatively, one can forward model into the observer-frame, and
and colours. All photometric catalogues are made publicly avail-
normalise by the sky area, to avoid any dependence on cosmological
able at https://camels.readthedocs.io/en/latest/data_
parameters.
access.html for the community to use.
In this study we focus on rest-frame, volume normalised distri-
Our findings are as follows:
bution functions, rather than observer frame fluxes. This is to better
understand the impact of various modelling assumptions on the rest- (i) We model luminosity functions and colour distributions in
frame properties, and better link to the underlying physics of each the GALEX FUV-NUV and SDSS ugriz photometric bands for the
model. We are also limited by the relatively small volume of the Swift-EAGLE, Illustris-TNG, Simba and Astrid galaxy formation
CAMELS simulations, which makes it more difficult to project into models from 𝑧 = 0.1 to 6, and find significant differences between
observer-space through the construction of a lightcone. their fiducial predictions, particularly in their optical and UV colours.
Finally, whilst observed luminosity functions from SDSS and (ii) Cosmological and astrophysical parameters significantly im-
GAMA are reasonably well covered in the intermediate magnitude pact the UV and optical luminosity functions and colours, with the
range over our 1P parameters (see Figures 5 – 11), it is clear that relative impact depending on the subgrid galaxy formation model.
we struggle to reproduce the detailed, complex colour distributions. In particular, we find that Ωm and 𝜎8 are both correlated with UV-
Whilst the broad trends are captured, some bins in colour are out optical colours, leading to redder galaxy distributions.
of distribution at this resolution. This is due to the known difficulty (iii) We build an SBI model using the LtU-ILI framework to
of replicating the detailed abundance and colour distributions in ob- predict the values of the cosmological and astrophysical parameters
servational space, even over this extended parameter space (see e.g. from our combined luminosity functions and colours, and achieve
Trayford et al. 2015). In future work we will investigate the impact of relatively accurate and precise constraints on Ωm , 𝜎8 , 𝐴SN1 & 𝐴SN2
varying more parameters in our forward model, such as the dust opti- across all galaxy formation models. The AGN parameters are poorly
cal depth and attenuation curve, to investigate whether this increases constrained due to limitations in the CAMELS simulation volumes
the compatibility of observed colours with the simulations. restricting the abundance of massive haloes.
For these reasons, we avoid performing inference directly on the (iv) We perform a feature importance analysis using the Illustris-
observed distribution functions in the rest-frame, and leave a detailed TNG model, and find that the tight constraints on Ωm are surprisingly
comparison with observations in the observer-frame to future work. driven by the UV-optical colour information. Colour information is
similarly important for 𝜎8 constraints.
(v) We investigate constraints from distribution functions at
6.3 Forward Modelling Limitations
higher redshifts, and find that 𝜎8 is best constrained at cosmic noon
As touched on briefly in Sections 3 and 2, there are a number of (𝑧 = 2), rather than 𝑧 = 0.1. The tightest constraints for all parameters
limitations in our modelling of the emission from galaxies, many of are obtained by combining information from multiple redshifts.
which are set by the fidelity of the fiducial CAMELS suites. Due to (vi) The constraints on 𝜎8 are surprising given the lack of any
the low spatial and mass resolution of the hydro simulations we do spatial or clustering information in our chosen features. We find
not resolve galaxy sizes, which precludes their inclusion as features. that enhanced clustering, parametrised by higher 𝜎8 , leads to earlier
We also therefore cannot include features such as emission profiles, forming and more metal enriched galaxies, which manifests through
or resolved maps of the emission in different bands. These features their star formation and metal enrichment histories, and subsequently
may contain crucial cosmological and astrophysical information, and their luminosities and colours.
will be regularly estimated in upcoming wide field surveys, such as (vii) As a test of robustness we apply a model trained on one
those from Euclid and Rubin. simulation to another, and find poor generalisability, as seen in many
The mass resolution in particular also places constraints on re- other CAMELS analyses. We attribute this to differences in the un-
MNRAS 000, 1–27 (2024)

derlying subgrid prescriptions in the galaxy formation models, but made use of the Spanish Virtual Observatory (https://svo.cab.inta-
also identify inflexibility in our forward model as a potential culprit. csic.es) project funded by MCIN/AEI/10.13039/501100011033/
through grant PID2020-112949GB-I00. TS acknowledges support
These results suggest that distributions of galaxy luminosities and by NSF grant AST-2421845 and NASA grants 80NSSC24K0089
colours across the UV-optical wavelength range contain not only a and 80NSSC24K0084. DAA acknowledges support by NSF grant
wealth of information on astrophysical processes, but also significant AST-2108944, NASA grant ATP23-0156, STScI grants JWST-
cosmological information content. Combined with more traditional GO-01712.009-A and JWST-AR-04357.001-A, Simons Foundation
sources of cosmological information, such as galaxy clustering statis- Award CCA-1018464, and Cottrell Scholar Award CS-CSA-2023-
tics, there is potential to use these properties to provide more stringent 028 by the Research Corporation for Science Advancement. AG is
constraints, and potentially break parameter degeneracies between grateful for support from the Simons Foundation.
probes. Calibration of subgrid models in hydrodynamic simulations We list here the roles and contributions of the authors accord-
of galaxy evolution is also an example where linking parameters to ing to the Contributor Roles Taxonomy (CRediT)7 . Christopher C.
data in a simulation-driven Bayesian approach has advantages (Elliott Lovell: Conceptualization, Data curation, Formal analysis, Investiga-
et al. 2021; Jo et al. 2023; Kugel et al. 2023); in this study we show tion, Methodology, Software, Visualization, Writing - original draft,
how this can be done directly in observer space, rather than relying Writing - review & editing. Tjitske Starkenburg: Conceptualization,
on inferred physical parameters obtained through SED fitting. Investigation, Validation, Writing - Review & Editing. Matthew Ho:
However, we have also shown the sensitivity of our inference to Methodology, Software, Investigation, Verification. Daniel Anglés-
the underlying galaxy formation model. Whilst we have discussed a Alcázar, Francisco Villaescusa-Navarro: Resources, Data Cura-
number of approaches for overcoming this in Section 5.4, in general tion, Writing - Review & Editing. Kartheik Iyer, Rachel Somerville,
improved models will help to reduce the domain shift when applying Laura Sommovigo: Conceptualization, Investigation, Writing - Re-
to real data. In addition, exploring larger parameter spaces with more view & Editing. William J. Roper: Methodology, Software, Writ-
physically motivated parameterisations of the astrophysics (e.g. the ing - Review & Editing. Austen Gabrielpillai, Alice E. Matthews:
28 parameter CAMELS SOBOL set; Ni et al. 2023) will help to cover Writing - review & editing.
a larger volume of the astrophysical and cosmological parameter
hypervolume.
In addition, the simple forward model for galaxy emission em-
ployed here can also be modified. The fidelity could be increased, DATA AVAILABILITY
for example by using radiative transfer approaches (Camps & Baes The full CAMELS public data repository is available at https:
2015; Narayanan et al. 2021) to self-consistently produce a wider //camels.readthedocs.io. This includes the photometric cata-
range of attenuation curves seen in the real Universe (Narayanan logues. Scripts for training and running the inference pipeline are
et al. 2018), or alternatively a larger range of parameters in the sim- made available at https://github.com/christopherlovell/
ple models employed here could be explored. This is one of the key camels_observational_catalogues. Download and installa-
development goals of synthesizer (Lovell et al. in prep., Roper et al. tion instructions for Synthesizer are available at https://
in prep.): to allow a rapid exploration of uncertain parameters in the flaresimulations.github.io/synthesizer/. Download and
forward model. This is still a huge data challenge, given the size of a installation instruction for the LtU-ILI package are provided at
simulation suite such as CAMELS (>200 million galaxies across all https://github.com/maho3/ltu-ili/tree/main.
suites and snapshots).
As described in Section 6.2, there are a number of difficulties to
comparing directly with observed data in an SBI framework. In future REFERENCES
work we will overcome a number of these limitations, at which point
the method will be directly applicable to a number of future surveys, Anglés-Alcázar D., Faucher-Giguère C.-A., Kereš D., Hopkins P. F., Quataert
such as Euclid (Euclid-Collaboration et al. 2024) and LSST (Ivezić E., Murray N., 2017, MNRAS, 470, 4698
et al. 2019), containing billions of sources. Appleby S., Davé R., Kraljic K., Anglés-Alcázar D., Narayanan D., 2020,
MNRAS, 494, 6053
Bellstedt S., Robotham A. S. G., 2024, ProGeny II: the impact of libraries
and model configurations on inferred galaxy properties in SED fit-
ting, doi:10.48550/arXiv.2410.17698, http://arxiv.org/abs/2410.
ACKNOWLEDGEMENTS 17698
Bernstein R. A., Freedman W. L., Madore B. F., 2002, ApJ, 571, 56
We thank Viraj Pandya for providing the GAMA catalogues used
Bird S., Ni Y., Di Matteo T., Croft R., Feng Y., Chen N., 2022, MNRAS, 512,
to construct the observed rest-frame colour distributions. We also 3703
thank Mat Page, Sotiria Fotopoulou, Shy Genel, Yongseok Jo, Borrow J., Anglés-Alcázar D., Davé R., 2020, MNRAS, 491, 6102
ChangHoon Hahn, Greg Bryan, Rachel Cochrane and Stephen Borrow J., Schaller M., Bower R. G., Schaye J., 2022, MNRAS, 511, 2367
Wilkins for useful discussions. This work was supported by the Si- Bose B., et al., 2021, MNRAS, 508, 2479
mons Collaboration on “Learning the Universe”. This work used Bravo M., Lagos C. d. P., Robotham A. S. G., Bellstedt S., Obreschkow D.,
the DiRAC Memory Intensive service (Cosma7) at Durham Uni- 2020, MNRAS, 497, 3026
versity, managed by the Institute for Computational Cosmology Bruzual G., Charlot S., 2003, MNRAS, 344, 1000
on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). Budavári T., et al., 2005, ApJ, 619, L31
The DiRAC service at Durham was funded by BEIS, UKRI and Bulichi T.-E., Davé R., Kraljic K., 2024, MNRAS, 529, 2595
Bullock J. S., Boylan-Kolchin M., 2017, ARAA, 55, 343
STFC capital funding, Durham University and STFC operations
Butler Contreras A., Lau E. T., Oppenheimer B. D., Bogd{\’a}n \., Tillman
grants. DiRAC is part of the UKRI Digital Research Infrastruc- M., Nagai D., Kovács O. E., Burkhart B., 2023, MNRAS, 519, 2251
ture. The data used in this work were, in part, hosted on equip-
ment supported by the Scientific Computing Core at the Flat-
iron Institute, a division of the Simons Foundation. This research 7 https://credit.niso.org/
MNRAS 000, 1–27 (2024)

Byler N., Dalcanton J. J., Conroy C., Johnson B. D., 2017, ApJ, 840, 44 physical Parameter Inference from Stacked Galaxy Cluster Profiles Us-
Byrne C. M., Stanway E. R., Eldridge J. J., McSwiney L., Townsend O. T., ing CAMELS-zoomGZ, doi:10.48550/arXiv.2410.10942, https://ui.
2022, MNRAS adsabs.harvard.edu/abs/2024arXiv241010942H
Calzetti D., 2001, New Astronomy Reviews, 45, 601 Ho M., et al., 2024, OJA, 7, 54
Camps P., Baes M., 2015, A&C, 9, 20 Hogg D. W., Baldry I. K., Blanton M. R., Eisenstein D. J., 2002, ArXiv
Carleo G., Cirac I., Cranmer K., Daudet L., Schuld M., Tishby N., Vogt- Astrophysics e-prints, pp arXiv:astro–ph/0210394
Maranto L., Zdeborová L., 2019, Rev. Mod. Phys., 91, 045002 Hopkins P. F., 2015, MNRAS, 450, 53
Carrilho P., Carrion K., Bose B., Pourtsidou A., Hidalgo J. C., Lombriser L., Ivezić \., et al., 2019, ApJ, 873, 111
Baldi M., 2022, MNRAS, 512, 3691 Iyer K. G., et al., 2020, MNRAS
Chabrier G., 2003, PASP, 115, 763 Jeffrey N., Alsing J., Lanusse F., 2021, MNRAS, 501, 954
Charlot S., Fall S. M., 2000, ApJ, 539, 718 Jo Y., et al., 2023, ApJ, 944, 67
Chatzikos M., et al., 2023, RMAA, 59, 327 Kugel R., et al., 2023, MNRAS, 526, 6103
Chawak C., Villaescusa-Navarro F., Echeverri-Rojas N., Ni Y., Hahn C., Lagos C. d. P., et al., 2019, MNRAS, 489, 4196
Anglés-Alcázar D., 2024, ApJ, 969, 105 Laughlin G., Bodenheimer P., Adams F. C., 1997, ApJ, 482, 420
Choustikov N., et al., 2024, MNRAS, 532, 2463 Le-Khac P. H., Healy G., Smeaton A. F., 2020, IEEE Access
Christensen N., Meyer R., Knox L., Luey B., 2001, Class. Quantum Grav., Leclercq F., Heavens A., 2021, MNRAS: Letters, 506, L85
18, 2677 Lee J. C., et al., 2009, ApJ, 706, 599
Conroy C., 2013, ARAA, 51, 393 Lemos P., Cranmer M., Abidi M., Hahn C., Eickenberg M., Massara E., Yallup
Crain R. A., van de Voort F., 2023, ARAA, 61, 473 D., Ho S., 2023, ML: ST, 4, 01LT01
Crain R. A., et al., 2015, MNRAS, 450, 1937 Li A., Liang S. L., Kann D. A., Wei D. M., Klose S., Wang Y. J., 2008, ApJ,
Cranmer K., Brehmer J., Louppe G., 2020, PNAS, 117, 30055 685, 1046
Davis M., Efstathiou G., Frenk C. S., White S. D. M., 1985, ApJ, 292, 371 Li Q., Narayanan D., Davé R., 2019, MNRAS, 490, 1425
Davé R., Anglés-Alcázar D., Narayanan D., Li Q., Rafieferantsoa M. H., Liske J., et al., 2015, MNRAS, 452, 2087
Appleby S., 2019, MNRAS, 486, 2827 Loveday J., et al., 2012, MNRAS, 420, 1239
Delgado A. M., et al., 2023, MNRAS, 526, 5306 Lovell C. C., Acquaviva V., Thomas P. A., Iyer K. G., Gawiser E., Wilkins
Dickey C. M., et al., 2021, ApJ, 915, 53 S. M., 2019, MNRAS, 490, 5503
Donnari M., et al., 2019, MNRAS, 485, 4817 Lovell C. C., et al., 2023, ICML 2023 ML4Astrophysics
Donnari M., Pillepich A., Nelson D., Marinacci F., Vogelsberger M., Hern- Madau P., Dickinson M., 2014, ARAA, 52, 415
quist L., 2021, MNRAS, 506, 4760 Marin J.-M., Pudlo P., Robert C. P., Ryder R. J., 2012, Stat Comput, 22, 1167
Dopita M. A., Sutherland R. S., 1996, ApJSS, 102, 161 McClintock T., et al., 2019, ApJ, 872, 53
Draine B. T., 2003, ARAA, 41, 241 Mead A. J., Brieden S., Tröster T., Heymans C., 2021, MNRAS, 502, 1401
Durkan C., Bekasov A., Murray I., Papamakarios G., 2019, in 33rd Con- Narayanan D., Conroy C., Dave R., Johnson B., Popping G., 2018, preprint,
ference on Neural Information Processing Systems (NeurIPS 2019). 1805, arXiv:1805.06905
https://ui.adsabs.harvard.edu/abs/2019arXiv190604032D Narayanan D., et al., 2021, ApJS, 252, 12
Echeverri-Rojas N., et al., 2023, ApJ, 954, 125 Nelson D., et al., 2018, MNRAS, 475, 624
Eldridge J. J., Stanway E. R., Xiao L., McClelland L. A. S., Taylor G., Ng Ni Y., et al., 2022, MNRAS, 513, 670
M., Greis S. M. L., Bray J. C., 2017, PASA, 34, e058 Ni Y., et al., 2023, arXiv.2304.02096
Elliott E. J., Baugh C. M., Lacey C. G., 2021, MNRAS, 506, 4011 Nicola A., et al., 2022, JCAP, 2022, 046
Euclid-Collaboration et al., 2024, arXiv Pallottini A., et al., 2022, MNRAS, 513, 5621
Fischera J., Dopita M. A., Sutherland R. S., 2003, ApJ, 599, L21 Pandya V., et al., 2017, MNRAS, 472, 2054
Fortuni F., et al., 2023, A&A, 677, A102 Papamakarios G., Murray I., 2016, in Advances in Neural In-
Friedman R., Hassan S., 2022, in Machine Learning and the Physical Sciences formation Processing Systems. Curran Associates, Inc.,
workshop, NeurIPS 2022. https://ui.adsabs.harvard.edu/abs/ https://proceedings.neurips.cc/paper/2016/hash/
2022arXiv221112724F 6aca97005c68f1206823815f66102863-Abstract.html
Furlong M., et al., 2015, MNRAS, 450, 4486 Pei Y. C., 1992, ApJ, 395, 130
Gabrielpillai A., Somerville R. S., Genel S., Rodriguez-Gomez V., Pandya Perez L. A., Genel S., Villaescusa-Navarro F., Somerville R. S., Gabrielpillai
V., Yung L. Y. A., Hernquist L., 2022, MNRAS, 517, 6091 A., Anglés-Alcázar D., Wandelt B. D., Yung L. Y. A., 2023, ApJ, 954,
Gebek A., Trčka A., Baes M., Martorano M., Pillepich A., Kapoor A. U., 11
Nersesian A., van der Wel A., 2024, MNRAS, 531, 3839 Pillepich A., et al., 2018, MNRAS, 473, 4077
Gebhardt M., et al., 2024, MNRAS, 529, 4896 Rodrigo C., Solano E., 2020, The SVO Filter Profile Service. https://ui.
Genel S., et al., 2014, MNRAS, 445, 175 adsabs.harvard.edu/abs/2020sea..confE.182R
Genel S., et al., 2019, ApJ, 871, 21 Rodrigo C., Solano E., Bayo A., 2012, Technical report, SVO Filter Pro-
Girardi L., Bressan A., Bertelli G., Chiosi C., 2000, Astron. Astrophys. Suppl. file Service Version 1.0, https://ui.adsabs.harvard.edu/abs/
Ser., 141, 371 2012ivoa.rept.1015R, doi:10.5479/ADS/bib/2012ivoa.rept.1015R. ,
Girardi L., Bertelli G., Bressan A., Chiosi C., Groenewegen M. a. T., Marigo https://ui.adsabs.harvard.edu/abs/2012ivoa.rept.1015R
P., Salasnich B., Weiss A., 2002, A&A, 391, 195 Rodríguez Montero F., Davé R., Wild V., Anglés-Alcázar D., Narayanan D.,
Greenberg D., Nonnenmacher M., Macke J., 2019, in Proceedings of the 36th 2019, MNRAS, 490, 2139
International Conference on Machine Learning. PMLR, pp 2404–2414, Roncoli A., Ćiprĳanović A., Voetberg M., Villaescusa-Navarro F., Nord B.,
https://proceedings.mlr.press/v97/greenberg19a.html 2023, Machine Learning and the Physical Sciences Workshop, NeurIPS
Hadzhiyska B., Bose S., Eisenstein D., Hernquist L., Spergel D. N., 2020, 2023
MNRAS, 493, 5506 Salim S., Narayanan D., 2020, ARAA, 58, 529
Hahn O., Abel T., 2011, MNRAS, 415, 2101 Schaller M., Schaye J., Kugel R., Broxterman J. C., Daalen M. P. v.,
Hahn C., et al., 2022, ApJ, 926, 122 2024a, The FLAMINGO project: Baryon effects on the matter power
Hahn C., Villaescusa-Navarro F., Melchior P., Teyssier R., 2024, ApJL, 966, spectrum, doi:10.48550/arXiv.2410.17109, http://arxiv.org/abs/
L18 2410.17109
Hassan S., et al., 2022, ApJ, 937, 83 Schaller M., et al., 2024b, MNRAS, 530, 2378
Hernández-Martínez E., Genel S., Villaescusa-Navarro F., Steinwandel U. P., Schaye J., et al., 2015, MNRAS, 446, 521
Lee M. E., Lau E. T., Spergel D. N., 2024, Cosmological and Astro- Shao H., et al., 2022, Robust field-level inference with dark matter ha-
MNRAS 000, 1–27 (2024)

los, doi:10.48550/arXiv.2209.06843, http://arxiv.org/abs/2209.
06843
Smith M. J., Geach J. E., 2023, RSOS, 10, 221454
Somerville R. S., Davé R., 2015, ARAA, 53, 51
Springel V., 2010, MNRAS, 401, 791
Springel V., White S. D. M., Tormen G., Kauffmann G., 2001, MNRAS, 328,
726
Stanway E. R., Eldridge J. J., 2018, MNRAS, 479, 75
Stanway E. R., Eldridge J. J., Becker G. D., 2016, MNRAS, 456, 485
Steinacker J., Baes M., Gordon K. D., 2013, ARAA, 51, 63
Talts S., Betancourt M., Simpson D., Vehtari A., Gelman A., 2018, arXiv
Tessore N., Loureiro A., Joachimi B., von Wietersheim-Kramsta M., Jeffrey
N., 2023, OJA, 6, 11
Thiele L., et al., 2022, Phys. Rev. D, 105, 083505
Tillman M. T., et al., 2023, AJ, 166, 228
Tortorelli L., McCullough J., Gruen D., 2024, The impact of stellar popula-
tion synthesis choices on forward-modelling-based redshift distribution
estimates, doi:10.48550/arXiv.2405.06009, http://arxiv.org/abs/
2405.06009
Trayford J. W., et al., 2015, MNRAS, 452, 2879
Trayford J. W., et al., 2017, MNRAS, 470, 771
Trayford J. W., Lagos C. d. P., Robotham A. S. G., Obreschkow D., 2020,
MNRAS, 491, 3937
Trčka A., et al., 2022, MNRAS, 516, 3728
Vĳayan A. P., Lovell C. C., Wilkins S. M., Thomas P. A., Barnes D. J.,
Irodotou D., Kuusisto J., Roper W. J., 2021, MNRAS, 501, 3289 Figure 1. TARP coverage for the fiducial Illustris-TNG model, including dust
Villaescusa-Navarro F., et al., 2020, ApJS, 250, 2 attenuated luminosity functions and colours, at 𝑧 = 0.1.
Villaescusa-Navarro F., et al., 2021a, arXiv:2109.09747
Villaescusa-Navarro F., et al., 2021b, ApJ, 915, 71
Villaescusa-Navarro F., et al., 2022, ApJ, 929, 132
Villaescusa-Navarro F., et al., 2023, ApJSS, 265, 54
Villanueva-Domingo P., Villaescusa-Navarro F., 2022, ApJ, 937, 115
Vogelsberger M., Marinacci F., Torrey P., Puchwein E., 2020, Nature Reviews
Physics, 2, 42
Weinberger R., et al., 2018, MNRAS, 479, 4056
Weinberger R., Springel V., Pakmor R., 2019, arXiv e-prints, 1909,
arXiv:1909.04667
Wilkins S. M., Feng Y., Di-Matteo T., Croft R., Stanway E. R., Bunker A.,
Waters D., Lovell C., 2016, MNRAS, 460, 3170
Wilkins S. M., et al., 2020, MNRAS, 493, 6079
Willmer C. N. A., 2018, ApJS, 236, 47
Zheng Y., Dave R., Wild V., Montero F. R., 2022, MNRAS, 513, 27
de Santi N. S. M., et al., 2023a, Field-level simulation-based infer-
ence with galaxy catalogs: the impact of systematic effects,
doi:10.48550/arXiv.2310.15234, http://arxiv.org/abs/2310.
15234
de Santi N. S. M., et al., 2023b, ApJ, 952, 69
von Wietersheim-Kramsta M., Lin K., Tessore N., Joachimi B., Loureiro A.,
Reischke R., Wright A. H., 2024, KiDS-SBI: Simulation-Based Inference
Analysis of KiDS-1000 Cosmic Shear, doi:10.48550/arXiv.2404.15402,
https://ui.adsabs.harvard.edu/abs/2024arXiv240415402V
Ćiprĳanović A., Kafkes D., Jenkins S., Downey K., Perdue G. N.,
Madireddy S., Johnston T., Nord B., 2020, in Workshop at
the 34th Conference on Neural Information Processing Systems
(NeurIPS). , doi:10.48550/arXiv.2011.03591, https://ui.adsabs.
harvard.edu/abs/2020arXiv201103591C
Ćiprĳanović A., Lewis A., Pedro K., Madireddy S., Nord B., Perdue G. N., Figure 2. Combined marginal posterior distribution of Ωm , 𝜎8 , 𝐴SN1 and
Wild S. M., 2022, in Workshop at the 36th conference on Neural Infor- 𝐴SN2 from all test set objects in the Illustris-TNG suite for the intrinsic
mation Processing Systems (NeurIPS). , doi:10.48550/arXiv.2211.00677, (orange) and dust attenuated (blue) LFs and colours.
https://ui.adsabs.harvard.edu/abs/2022arXiv221100677C
in filters from the GALEX (𝐹𝑈𝑉 and 𝑁𝑈𝑉), GAMA (𝑢𝑔𝑟𝑖𝑧𝑌 𝐽𝐻𝐾),
Johnson (𝑈𝐵𝑉 𝐽), HST ACS (F435W, F606W, F775W, F814W, and
APPENDIX A: PHOTOMETRY DATABASE
F850LP), HST WFC3 (F098M, F105W, F110W, F125W, F140W
The photometric catalogues produced in this work are avail- and F160W), and JWST NIRCam (F070W, F090W, F115W, F150W,
able at https://camels.readthedocs.io/en/latest/data_ F200W, F277W, F356W and F444W) instruments. These are pro-
access.html. We provide AB rest- and observer-frame magnitudes vided for both the BPASS and BC03 SPS models, and for intrinsic
MNRAS 000, 1–27 (2024)

and dust attenuated emission. Full details on the forward modelling
pipeline are provided in Section 3.
APPENDIX B: COVERAGE TESTS AND CALIBRATION

We carried out a number of tests of the model coverage for each
of our trained SBI models, in order to calibrate our model (Talts
et al. 2018). In each case we used the 10% of reserved simulation
in the test set. We produced and considered P-P plots, to compare
the CDF of the probability integral transform to that of a uniform
random variable, as well as the posterior predictions directly, to
assess the accuracy and precision of each marginal posterior. We also
used approximate methods to test our posterior coverage, including
the Tests of Accuracy with Random Points (Lemos et al. 2023).
TARP evaluates distances between samples of the posterior, the true
value, and a random point, and uses these to estimate the expected
coverage probability; this approach has been shown to provide a
robust test of the optimality of the posterior estimator. An example
for the Illustris-TNG model (including LFs and colours) is shown in
Figure 1, showing excellent coverage and little to no bias. Similar
results are achieved across all of our trained models.
APPENDIX C: CONSTRAINTS FROM THE INTRINSIC

EMISSION
The dust model described in Section 3.2 does not depend on galaxy
properties explicitly. However, it does attenuate young stars (< 10
Myr) more than old stars, and as such it depends on the shape of
the star formation history of each galaxy around this pivot point.
Galaxies with younger star formation histories will experience more
attenuation; this may then act either as a degeneracy with the underly-
ing cosmological and astrophysical parameters, or provide additional
information. To test this, we present in Figure 2 the stacked marginal
posterior distributions using attenuated or intrinsic LFs and colours
in the Illustris-TNG model. For the cosmological parameters there
is no difference in the RMSE, however for the supernovae feedback
parameters the constraints are slightly tighter in the model using
attenuated emission compared to the intrinsic emission. This effect,
while small, suggests that the differential attenuation of old and young
stars imprints additional information from the feedback parameters
on the emission, from which the model can learn.
APPENDIX D: FEATURE IMPORTANCE ACROSS

ADDITIONAL MODELS
Figures B1, D1 and D2 show the stacked binned residuals of the
marginal posteriors for the cosmological and supernovae feedback
parameters in the Simba, Swift-EAGLE and Astrid simulations.
Similarly to the trends seen in Figure 15 for Illustris-TNG, we
find that for all simulations it is the colours that provide the most
information on Ωm . We also see that the information at 𝑧 = 2 is most
informative for constraints on 𝜎8 , as compared to lower redshift
luminosity functions and colours.
This paper has been typeset from a TEX/LATEX file prepared by the author.
MNRAS 000, 1–27 (2024)

Figure B1. The same as Figure 15, but showing the combined posteriors for the Simba simulations.
Figure D1. The same as Figures 15 and B1, but showing the combined posteriors for the Swift-EAGLE simulations.
MNRAS 000, 1–27 (2024)

Figure D2. The same as Figures 15 and B1, but showing the combined posteriors for the Astrid simulations.
MNRAS 000, 1–27 (2024)

Learning The Universe: Cosmological and Astrophysical Parameter Inference With Galaxy Luminosity Functions and Colours

Uploaded by

Document Informationclick to expand document information

Document Informationclick to expand document information

Copyright:

Available Formats

Learning The Universe: Cosmological and Astrophysical Parameter Inference With Galaxy Luminosity Functions and Colours

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Learning The Universe: Cosmological and Astrophysical Parameter Inference With Galaxy Luminosity Functions and Colours

Uploaded by

Copyright:

Available Formats

MNRAS 000, 1–27 (2024) Preprint 22 November 2024 Compiled using MNRAS LATEX style file v3.

Learning the Universe: Cosmological and Astrophysical Parameter

Accepted XXX. Received YYY; in original form ZZZ

1 INTRODUCTION tion of these methods is in Bayesian inference, wherein one estimates

© 2024 The Authors

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

Simulation 𝐴SN1 𝐴SN2 𝐴AGN2 𝐴AGN2

3.1 Stellar population modelling

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

population at intermediate redshifts, before settling to a Schechter- 4.2 1P set distributions

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

APPENDIX B: COVERAGE TESTS AND CALIBRATION

APPENDIX C: CONSTRAINTS FROM THE INTRINSIC

APPENDIX D: FEATURE IMPORTANCE ACROSS

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

MNRAS 000, 1–27 (2024)

You might also like